Merge remote-tracking branch 'origin/rustify' into worktree-agent-a27094a7db775552a

# Conflicts: # docs/IRIS.md
docs: P0's iris half done -- bench feature, emulator smoke run, APK
2026-09-05 21:37:04 -04:00 · 2026-09-05 21:35:51 -04:00 · 2026-09-05 21:35:44 -04:00 · 2026-09-05 21:35:33 -04:00 · 2026-09-05 21:18:30 -04:00 · 2026-09-05 21:10:19 -04:00
93 changed files with 14438 additions and 784 deletions
@@ -0,0 +1,244 @@
+---
+name: ai-app-rigs
+description: ai-app's test rigs, harness scripts and reference measurements - ui-sandbox.sh, debug-transcript.sh, transcript-bench.sh, stream-bench.sh, trace-draw.sh, the /usage fixture vocabulary, the fake CLI, the rule that no UI-driving script may tap a coordinate, how to test llama.cpp and ssh on this machine, how importing behaves, and the scroll/stream/explorer numbers not worth re-measuring. Read before running or writing a benchmark, driving the app's UI from a script, exercising the session lifecycle, testing a llama or remote session, or touching the import screen.
+---
+
+# ai-app: rigs, harnesses and measurements
+
+Moved out of `AGENTS.md` on 2026-09-04 so it is read when it is relevant
+rather than sent with every request in this repo -- it was 12 KB of the 35 KB
+that file cost on every one. Unchanged in the move, and still the only copy.
+
+## The rigs
+
+Each exists because something was invisible without it.
+
+- **`app/ui-sandbox.sh`** — a second `ai-server` with its own `$HOME`, config
+  and data directory, holding eight invented Claude Code transcripts and a
+  `claude` that is two lines of shell. **That isolation is the point**: the
+  import screen lists whatever is in `~/.claude/projects`, which in this VM is
+  real agent transcripts, so exercising *delete* against the ordinary server
+  deletes somebody's conversation and exercising *import* starts a real
+  `--resume` on the owner's account.
+  Its port and root derive from the checkout's name, so two checkouts'
+  sandboxes cannot reach each other, and its token is generated once into
+  `~/.config/ai-app/sandbox-token` and carried across restarts along with any
+  the enrolment flow appended — so the emulator app is enrolled **once** (the
+  start banner prints the command) and stays enrolled. It shares the real TLS
+  certificates, because the installed APK pins that CA.
+  Driving verbs, so none of this is re-derived per session:
+  `./ui-sandbox.sh spawn [title]` (an echo session, prints its id),
+  `./ui-sandbox.sh send SID text|@file`, and
+  `./ui-sandbox.sh api /path [curl args]`.
+  `./ui-sandbox.sh keep` restarts the server without wiping the sessions and
+  enrolment already there — for when the fixture under test was expensive to
+  build; plain `start` wipes them, which is right for the list-screen
+  fixtures and wrong for that.
+  It passes `--delay` by default, and `AI_SANDBOX_BIG_MB` puts one large
+  transcript among the small ones while `AI_SANDBOX_SPAWN_DELAY` makes the
+  fake CLI slow to start. Both exist because operations that finish in
+  milliseconds have states on the way that nothing can observe, and an
+  unobservable state is one where broken and working look identical.
+  It also builds a fixture tree at the sandbox home's `~/files` for the
+  explorer, holding the states otherwise only reachable by finding a real
+  machine in one: an empty directory, a name with a tab and one with an
+  apostrophe, a binary file, one over `FILE_LIMIT`, one `chmod 000`, a
+  symlink to a directory and a broken one, a source file per language, and
+  the three sizes the limits were measured against (`edit-32k.rs`,
+  `edit-128k.rs`, `big-source.rs`). Point a session at it with
+  `./ui-sandbox.sh api /sessions/<id>/cwd -X POST -H 'content-type: application/json' -d '{"cwd":"~/files"}'`.
+  The explorer's 409 is produced by editing the file on the machine
+  (`printf … > file`) between pressing the pencil and pressing save.
+- **`app/debug-transcript.sh`** — a real conversation on the emulator. The
+  echo driver is the right rig for most things and the wrong one for anything
+  whose cost scales with what was actually written: a real reply is longer,
+  is real markdown, and carries tool calls whose input and output are
+  kilobytes. Two faults were invisible until a real transcript was loaded — a
+  page of history landing mid-fling threw the reader back to the newest end,
+  and parsing one real reply took 51ms against 4.6ms for a synthetic one.
+  `-b` takes the biggest conversation on the machine rather than the newest,
+  which is what a scrolling test wants; `--stop` takes it down.
+  It copies the transcript into `/tmp` and gives the server a `HOME` of its
+  own, so the import can only see the copy — importing spawns `claude
+  --resume`, and against the real file that is a second CLI writing to a
+  conversation somebody may still be in. **A transcript never goes in this
+  repository**: they hold whatever was said, read and written in that
+  session, and `~/repos` is shared with the host besides.
+- **`/usage` in an echo session puts up an invented meter**, which is how the
+  rate-limit screens' states are reached without spending quota: `/usage 42`,
+  `/usage 95 20` (minutes left), `/usage 42 never` (the between-blocks window
+  with no reset time), `/usage 42 unreadable`, `/usage notloggedin`,
+  `/usage unreachable`, `/usage failed`, `/usage off`. The vocabulary is
+  `usage::Fixture`'s, since those are its states. With none set an echo
+  session meters nothing, which is the ordinary case and draws no bar.
+- **A fake CLI exercises the process lifecycle without a token.** Point a
+  `claude_cli` provider's `command` at a two-line script — `#!/bin/sh` and
+  `cat > /dev/null` — and it behaves the way the lifecycle code cares about:
+  it holds the fifo open, records a real pid, writes nothing, and dies on a
+  signal. So adopt, stop, restart and start are all drivable without a real
+  `--resume` and without spending a turn on somebody's account. Reach for
+  this when what is under test is *whether a process is running*, and for
+  `debug-transcript.sh` when it is *what the transcript draws*.
+- **`app/transcript-bench.sh`** is the standard scroll measurement: it opens
+  the first session (or `-k` keeps the current screen), scrolls a fixed
+  gesture loop, and prints the app's render report — the same one the in-app
+  copy button produces, whose `on screen:` line names what the viewport was
+  holding. Compare two runs with the same gestures; the emulator's absolute
+  frame times transfer nothing, the report's accounting does. Run it either
+  side of any change under `Markdown*.kt`, `Transcript*.kt` or
+  `SessionScreen.kt`'s list, and put the report in the commit. The numbers
+  that move first are the worst `record: one block`, the reparse mean while
+  streaming, and the draw phase's accounting line.
+- **`app/stream-bench.sh [-k] FILE`** is that measurement for a reply still
+  arriving. It taps "Jump to latest" so the list is pinned to the newest end,
+  resets the report, sends FILE, waits for the transcript to stop growing,
+  and prints. Both of those are corrections to a first version that measured
+  nothing: a transcript parked further back never redraws while a reply
+  streams into it, and a session is idle at *both* ends of a turn, so polling
+  for idle answers before the turn has started.
+- **`app/trace-draw.sh`** names what a scrolling frame spends inside the
+  framework, from `atrace` text output with no trace processor needed. It is
+  how the cost of a layout node per link was attributed to the framework
+  rather than guessed at.
+
+### Driving the UI
+
+**No script that drives this app's UI presses a coordinate.** Every control
+is found by the name it already carries for assistive technology —
+`ui-trace record --do "tap 'Session settings'"` — which resolves the label
+against the screen at the moment of the gesture and fails the whole run when
+it is not there. `app/bench-lib.sh` is what the bench scripts share for it. A
+coordinate is a position measured once by hand, and anything that moves the
+control makes the tap land on whatever now sits there — the bench then
+reports a number that was never measured, which reads exactly like a result.
+Both bench scripts pressed the render report at `tap 723 205` until that
+button moved into the session settings dialog on 2026-09-03. The check that
+none has crept back:
+
+    grep -n "tap [0-9]" app/*.sh
+
+Swipes are still coordinates, deliberately: a gesture across a scrolling area
+is a distance rather than a control.
+
+**Two traps in the emulator bench loop**, each of which cost a run.
+`adb shell pm clear` removes the enrolment and the notification permission
+along with the saved anchors, so the next run measures a permission dialog —
+re-enrol with the command `ui-sandbox.sh` prints, and
+`pm grant … POST_NOTIFICATIONS`. And a saved scroll anchor is per session id,
+so the only way two builds start a scroll from the same place is a *fresh
+session for each*.
+
+**The emulator is `~/repos/emulator-tools`' business, not this repo's.**
+`emu up` creates and boots the AVD named after this checkout — whatever `emu
+name` prints, never a name typed out here, since this file is the same in
+every clone. `run-android.sh` is that plus a build and an install. The `adb`
+on `PATH` after sourcing `android-env.sh` is that repo's wrapper, which fills
+in `-s` from the same rule. Gradle does not go through it, so a Gradle init
+script from `emulator-tools` runs `emu check` before `installDebug`,
+`uninstallDebug` and `connectedAndroidTest` and fails rather than fanning out
+to every attached device; when it refuses, say which device you mean at the
+moment you use it — `ANDROID_SERIAL=$(emu serial) ./gradlew …`.
+
+### Testing llama.cpp and ssh here
+
+**Both are set up here as of 2026-09-04** and need nothing typed. The
+prebuilt CPU llama.cpp lives outside the repo at `~/.local/opt/llama.cpp`
+(the 15 MB `ubuntu-x64` release asset) and is symlinked as
+`/usr/local/bin/llama-server`, which is what makes **discovery find it over
+ssh**: `~/.local/bin` is not on the PATH a non-interactive ssh session gets.
+It resolves its own libraries through `$ORIGIN`, so no `LD_LIBRARY_PATH` is
+needed. One model is downloaded — `unsloth/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.gguf`,
+639 MB under `~/.local/share/ai-app/models` — and answers at usable speed on
+this VM's 8 cores. **Do not test with a 2-bit quant**: the
+IQ2_XXS of that model produces fluent nonsense, which reads exactly like a
+broken driver — `llama-cli` produces the same from the file directly, which
+is how to tell the two apart in a hurry.
+
+There is no second machine, so **ssh this VM to itself**. That is set up
+too: the key is `~/.config/ai-app/ssh-self` (its public half is in
+`~/.ssh/authorized_keys`, labelled removable), and the real config carries a
+setup called **"this vm over ssh"** — `bob@127.0.0.1` with that
+`identityFile` plus
+`options: ["StrictHostKeyChecking=no", "UserKnownHostsFile=/tmp/ai-app-known-hosts"]`
+so it touches nothing real — offering `claude-cli` and `llama-cpp`. It is the
+whole rig for "does a remote llama session work", since the far machine is
+this one and the model file is the same file. For a throwaway setup of your
+own, point a provider's `command` at something harmless like `/bin/echo`
+rather than at `claude`: the transport is what is under test, the process
+exiting immediately is the signal, and it costs no tokens. The remote login
+shell here is **fish**; the
+remote script and `ssh.rs`'s POSIX quoting happen to mean the same thing in
+both, but that is luck rather than design, and a shell that is neither is the
+thing to suspect first if a remote spawn ever mangles an argument.
+
+## Importing
+
+The import list reports each session's **size as well as its line count**,
+because the two disagree in the way that matters: these transcripts embed
+screenshots as base64, so one line can be a megabyte. On this machine a 69 MB
+session has 3,427 lines and a 44 MB one has 6,792 — nothing about a line
+count tells you what continuing a session will cost. Shown, not warned about;
+importing a large session is a choice somebody is entitled to make.
+
+**Never import a Claude Code session that is open in a terminal.** The app
+refuses it — see PLAN.md for the incident that made that a refusal rather
+than a warning.
+
+**One Claude Code session id can name two files, and the listing offers it
+once.** Resuming from a different working directory makes the CLI write a
+second transcript with the same id under that directory's project folder — an
+ordinary state of a machine, not corruption. Everything downstream addresses
+a session by id, and the phone keyed its list on it, so two rows sharing one
+**closed the app** on a Compose duplicate-key throw. `parse_listing` keeps
+the copy with the most lines, because the other is usually a few-hundred-byte
+stub and is often the *newer* of the two, so recency is the wrong key.
+Deleting removes every copy rather than the first, or the row came back after
+a delete that reported success. The phone's half is `uniqueItems`, which
+every list keyed on a server-chosen id goes through: a repeat there must
+never be able to close the app, whatever produced it.
+
+**Deleting a session offers to take the machine's own transcript with it** —
+`DELETE /sessions/{id}?deleteForeign=true`, behind a switch in the
+confirmation, and only where the driver keeps a record of its own
+(`keepsOwnTranscript`, which today means Claude Code). Off by default,
+because leaving that copy is what makes an ordinary delete recoverable — and
+the dialog's paragraph is rewritten when it is on rather than appended to,
+since the sentence promising the conversation "should still be there to
+import again" is exactly the one the switch makes false. The server deletes
+the machine's copy *first*, so a machine it cannot reach leaves the session
+where it was instead of half-deleted.
+
+## Measurements worth not re-taking
+
+- **What the transcript screen costs to scroll.** Taken 2026-08-30 on the GPU
+  emulator against a real imported transcript with the server at
+  `--delay 120`. Settled and flinging fast, both into fresh history and back
+  through rows already drawn: **5.2–5.9% janky frames, 99th percentile
+  29–32ms, 0–2 slow UI-thread frames.** The stock Settings app on the same
+  device is 3.3% and 38ms, so this is at the platform floor. The number that
+  is *not* at the floor is the first few seconds after opening a session,
+  where every row on the way is being composed for the first time; that is
+  inherent to a lazy list and it is why a measurement taken before the screen
+  settles reads three times worse. **Settle first, then reset `gfxinfo`.**
+- **The reset path is not reachable by reopening a session.** Measured
+  2026-09-04 against a session streaming at 20 events a second: reopening one
+  with an anchor 1,800 events back connects **87–119 events behind**, well
+  under `CATCH_UP_LIMIT`'s 200, because the restore is two requests — the
+  opening page, then one span covering the whole distance. To exercise the
+  reset at all you have to lower `CATCH_UP_LIMIT` in a throwaway build; at 5
+  the app takes the reset on a live connection, clears, refills and carries
+  on without reconnecting.
+- **The session screen's stream survives backgrounding here** — 20 seconds at
+  the launcher while 415 events were produced brought no reconnect at all,
+  which is not what the comment above that loop expects, and is most likely
+  this emulator being headless rather than the phone's behaviour.
+- **Reopening a cached session costs one request for one event** (the probe),
+  and scrolling the whole conversation back costs nothing more; a cold open
+  of the same 500-event session is two pages, 100 events. Measured
+  2026-09-04 on the emulator against the sandbox.
+- **Reading is cheap and editing is not.** The viewer handles a 1 MiB,
+  28,000-line file because it draws one row per line; the editor is one
+  `BasicTextField`, which costs two seconds a frame at 128 kB and stops the
+  app at 1 MiB, so `EDIT_LIMIT` caps it at 32 kB with the reason said on
+  screen. If you make the editor faster, that number is what to move.
+  EXPLORER.md's "What the measurements said" has the rest.
@@ -5,11 +5,14 @@ replacing the Claude app for daily use. Rust/Axum backend on the desktop,
 Kotlin/Compose Android app, WireGuard + pinned self-signed TLS + bearer token
 between them.

-**`PLAN.md` is the design source of truth** — every decision with its date,
-its rationale, and what was rejected. Read it before changing anything
+**`docs/PLAN.md` is the design source of truth** — every decision with its
+date, its rationale, and what was rejected. Read it before changing anything
 structural, and update it in place when a decision changes rather than
 letting this file and the plan become two versions of the truth. This file is
 the working notes layer: layout, commands, rigs, and things that have bitten.
+The design and working documents live under `docs/` — everything except this
+file and `CLAUDE.md`, which stay at the root because that is where Claude
+Code and other agent harnesses look for them.

 The central design point, worth not undoing by accident: **a session is a
 child process, translated into one common event model.** A new session type
@@ -22,7 +25,7 @@ Mirrors `../dev-updater` deliberately: same stack (axum 0.8 +
 axum-server/rustls, tokio, clap; Kotlin 2.4.x + Compose Multiplatform, single
 `:androidApp` module), same cert scheme, same registry pattern. Read
 dev-updater's `README.md` and `AGENTS.md` before diverging from them.
-Module-by-module intent is in PLAN.md's "Backend layout".
+Module-by-module intent is in `docs/PLAN.md`'s "Backend layout".

 - `server/` — the Rust backend (`ai-server`). `routes.rs`'s module doc
  comment is the HTTP table and the surface's source of truth.
@@ -40,16 +43,23 @@ Module-by-module intent is in PLAN.md's "Backend layout".
  projects version-locked to the commit this repo pins. What deliberately did
  **not** move is the API surface and the config *schema*: routes, drivers,
  sessions and setups are what makes this project itself.
- `EXPLORER.md` — the file explorer's design (`server/src/files.rs` and
-  `FilesScreen.kt` / `FileViewer.kt` / `FileEditor.kt`).
- `TRANSCRIPT_CACHE.md` — the phone's copy of what it has been sent. Read it
-  before touching `TranscriptCache.kt`, `TranscriptSource.kt`, or the opening
-  and stream effects in `SessionScreen.kt`.
- `TODO.md` — the working list.
- `RUST.md` — the plan for moving the app to Rust (on the `rustify`
-  branch of the `ai-app-2` clone): what has to be reproduced, the
-  framework decision, and the ordered experiments with their pass
-  conditions. Read it before touching anything under that branch.
+- `docs/` — every design and working document except this file and
+  `CLAUDE.md`:
+  - `docs/EXPLORER.md` — the file explorer's design (`server/src/files.rs`
+    and `FilesScreen.kt` / `FileViewer.kt` / `FileEditor.kt`).
+  - `docs/TRANSCRIPT_CACHE.md` — the phone's copy of what it has been sent.
+    Read it before touching `TranscriptCache.kt`, `TranscriptSource.kt`, or
+    the opening and stream effects in `SessionScreen.kt`.
+  - `docs/TODO.md` — the working list.
+  - `docs/RUST.md` — the plan for moving the app to Rust (on the `rustify`
+    branch of the `ai-app-2` clone): what has to be reproduced, the
+    framework decision, and the ordered experiments with their pass
+    conditions. Read it before touching anything under that branch.
+  - `docs/IRIS.md`, `docs/IRIS_TODO.md`, `docs/DECISIONS.md`,
+    `docs/LAYOUT.md`, `docs/TEXTURES.md`, `docs/CLIENT_CORE.md` — iris's
+    own public API log, working list, decisions log, layout/render design,
+    and texture-atlas design, and the client-core crate's design,
+    respectively.
 - `.dev-updater.ron` — what Dev Updater builds here: the server (run as
  `service: Managed(…)`, supervised by Dev Updater's own implementation
  rather than a script kept here) and the APK, in parallel. It points at
@@ -86,7 +96,11 @@ two icon buttons the same width without either being given one — and why
  :androidApp:compileDebugKotlin :androidApp:lintDebug
  :androidApp:testDebugUnitTest`. The unit tests are JVM-only and cover the
  syntax highlighter, the ANSI parser and the transcript cache — the app's
-  pure logic with no Android in it.
+  pure logic with no Android in it. Touching anything under `BenchFixture.kt`,
+  `BenchNetwork.kt`, `BenchRun.kt` or the `bench` build type also needs
+  `:androidApp:compileBenchKotlin :androidApp:lintBench` — a second build
+  type compiles separately and lint has caught real bugs debug alone never
+  would (see "Android Lint" below).
 - **Android Lint is not optional and is not run by a build.** It found a
  crash that had been shipping (`java.time` on a minSdk-24 app with
  desugaring off) and later a permission check that silently dropped every
@@ -146,6 +160,27 @@ two icon buttons the same width without either being given one — and why

 Each exists because something was invisible without it.

+- **The `bench` build type and `app/bench-fixture/`** exist for P0 (RUST.md
+  and DECISIONS.md's 2026-09-05 entries), the phone benchmark gate Iris
+  asked for before porting continues: a deterministic, checked-in synthetic
+  transcript (`app/bench-fixture/generate.py`, never a real one) that both
+  this app and iris open with no server, so a frame-time comparison
+  measures the renderer rather than the data. `./build-apk.sh bench` builds
+  it — own application id (`com.example.aiapp.bench`) and label ("AI
+  Sessions bench") so it installs beside a real enrollment rather than
+  replacing it. Opening it goes straight to a session screen holding the
+  fixture (no enrollment, no permission prompts) with a "Run benchmark"
+  control beside "Copy" in session settings: it drives the same scroll loop
+  and streaming phase `transcript-bench.sh`/`stream-bench.sh` drive over
+  `ui-trace`, but in-process, since a real phone has no usable system
+  tracing and no agent can drive one (this-machine-android's skill).
+  `BenchFixture.kt`/`BenchNetwork.kt` fake the backend by installing a
+  `URLStreamHandlerFactory` that answers `TranscriptSource`/`EventStream`'s
+  requests from an in-memory copy of the fixture instead of opening a
+  socket — so the fold, the paging and `uniqueItems` under test are the
+  screen's real ones, never a shortcut built just for this. The report
+  gains a `bench:` section (process CPU time, peak RSS, battery current) on
+  every build, empty except when `BenchRun.kt` filled it in.
 - **`app/ui-sandbox.sh`** — a second `ai-server` with its own `$HOME`, config
  and data directory, holding eight invented Claude Code transcripts and a
  `claude` that is two lines of shell. **That isolation is the point**: the
@@ -312,7 +347,7 @@ means here:
 ## Sessions outlive the backend

 Since 2026-08-29 a session's process is deliberately left running when
-`ai-server` stops, and adopted again when it starts. PLAN.md has the design;
+`ai-server` stops, and adopted again when it starts. docs/PLAN.md has the design;
 day to day:

 - **Stopping the server no longer stops the sessions.** After `pkill
@@ -345,7 +380,7 @@ count tells you what continuing a session will cost. Shown, not warned about;
 importing a large session is a choice somebody is entitled to make.

 **Never import a Claude Code session that is open in a terminal.** The app
-refuses it — see PLAN.md for the incident that made that a refusal rather
+refuses it — see docs/PLAN.md for the incident that made that a refusal rather
 than a warning.

 **One Claude Code session id can name two files, and the listing offers it
@@ -521,4 +556,4 @@ belongs in `~/.claude/TOOLCHAIN.md` or `~/.claude/MACHINE.md` instead.
  `BasicTextField`, which costs two seconds a frame at 128 kB and stops the
  app at 1 MiB, so `EDIT_LIMIT` caps it at 32 kB with the reason said on
  screen. If you make the editor faster, that number is what to move.
-  EXPLORER.md's "What the measurements said" has the rest.
+  docs/EXPLORER.md's "What the measurements said" has the rest.
@@ -1,30 +1,44 @@
-# Decisions taken for Iris to review
+# Decisions awaiting review

-Short list of design choices made by the design agent without asking, so
-they can be judged and reversed later. Detail lives in RUST.md (and IRIS.md
-for iris API changes); this file is only the summary. Newest first. Items
-marked **DEFERRED** are ones the agent chose not to decide alone.
+Choices made while working autonomously, for Bryan to keep or change. Each
+says what was picked and why; the detail is in the design doc it names.
+Delete an entry once it has been looked at.

-## 2026-09-05
+## Subagent views (2026-09-05, `SUBAGENTS.md`)

- **Touch drag on a transcript row follows Android's own rule**: a vertical
-  drag pans the list immediately; a stationary press held 500 ms starts a
-  text selection which further dragging extends; a horizontal drag while
-  something is already selected extends that selection without the wait.
-  One `DragArbiter` per list decides it (`iris/src/sense.rs`). Chosen over a
-  "text layer always wins" or "list always wins" rule because either loses
-  one of the two gestures a reader expects.
- **E4's desktop shape is a new `iris/desktop-app` crate**: a winit window
-  holding `transcript-ui`'s screen beside a session list, talking to a real
-  `ai-server` through `client-core`. It enrols by pasting the same
-  `aiapp://enroll?…` link a phone scans (`client-core::config::EnrolledServer`)
-  and keeps it owner-only under `$XDG_CONFIG_HOME/ai-app-desktop/`. The
-  pinned CA is a path given on the command line, not baked in. Chosen so
-  the phone and desktop share one enrolment format and no second one is
-  invented.
- **Order of remaining work**: finish the two in-flight pieces above, then
-  the transcript screen's Android integration and the `transcript-bench.sh`
-  comparison against Compose — the numbers the recommendation still lacks.
- **DEFERRED — whether to commit to iris over Masonry for `ai-app`.** Waits
-  on the bench numbers above; RUST.md's recommendation says what the
-  measurements must show.
+Made on my own judgement, limited blast radius:
+
+1. **A subagent is a transcript, not a session.** It has no process,
+   controls or settings; it is addressed as `/sessions/{id}/subagents/{sub}`
+   and stored under the session's directory, so deleting the session takes
+   it. Alternative rejected: registering it as a session of its own, which
+   would give it a card in the main list and a driver that can do nothing.
+2. **Read-only view is the session screen minus its controls**, rather than
+   a second, simpler transcript screen. Keeps paging, caching, selection
+   and rendering in one place. Cost: a `readOnly` mode threaded through
+   `SessionScreen`.
+3. **The list only carries a count.** Each session row says how many
+   subagents it has; their titles and statuses are fetched when the card is
+   expanded. Keeps `GET /sessions` from reading every subagent transcript.
+   Consequence: an expanded card's statuses refresh with the list, not live.
+4. **Expanded/collapsed is remembered per session on the phone**, not on
+   the server. Collapsed by default, per the transcript convention that new
+   things arrive collapsed.
+5. **Subagents of imported sessions are not shown.** The import path still
+   skips `isSidechain` records; the CLI's own `subagents/agent-*.jsonl` files
+   are not read. Only subagents run while this backend was watching exist.
+6. **Echo grows `/subagent [n]`** as the test rig, so nothing here needs a
+   paid turn to exercise.
+
+Deferred, because they reach further than this feature:
+
+- **Live status on the list.** Whether the session list should follow a
+  stream at all (it refreshes on demand today) decides whether subagent
+  status can ever be live there. Not changed.
+- **Nested subagents.** A subagent's own Task calls are shown as tool calls
+  in its transcript and are not given transcripts of their own. Supporting
+  that is the same mechanism one level down, but the UI would need nested
+  expanders.
+
+- **The subagent status row says "context unknown".** Nothing measures a
+  subagent's context; the row could leave it out rather than admit it.
@@ -0,0 +1,141 @@
+# Subagents
+
+A session's subagents -- the helpers a Claude Code session starts through its
+Task tool -- each get a transcript of their own, listed under the session's
+card and readable in the same transcript view the session has. Designed
+2026-09-05; the decisions Bryan has not yet reviewed are in `DECISIONS.md`.
+
+## What a subagent is here
+
+**A subagent is a second transcript owned by a session, in the same event
+model, with no process and no controls.** It is not a session: it cannot be
+messaged, stopped or started, and it has no setup, model or usage of its
+own. Everything it shares with a session -- the transcript file format, the
+paging routes, the SSE stream, the phone's cache and rendering -- is reused
+by addressing, not by copying.
+
+The CLI reports a subagent's messages on the parent's own stream-json
+output, each carrying `parent_tool_use_id` = the id of the Task `tool_use`
+that started it. Before this the translator dropped those lines
+(`subagent_events_are_not_duplicated_into_the_transcript`); now it routes
+them to that subagent's own translator and transcript. The parent's
+transcript still shows only the Task call itself.
+
+## Storage
+
+Under the session directory:
+
+```
+<session>/subagents/<tool_use_id>/meta.json        {title, created}
+<session>/subagents/<tool_use_id>/transcript.jsonl same SeqEvent lines as the session's
+```
+
+The id is the Task tool_use id (`toolu_…`), which is unique, stable across a
+backend restart, and already the key everything on the parent side uses.
+Only ids matching `[A-Za-z0-9_-]+` are ever created or looked up, since the
+id becomes a path.
+
+The transcript's sequence numbers are its own, starting at 1. `Transcript`,
+`read_window`, `catch_up` and `read_after` work on it unchanged.
+
+Its path out: deleting the session deletes its directory, subagents included.
+There is no separate delete.
+
+## Lifecycle, as events in the subagent's transcript
+
+1. Created on the first child line for an unseen parent id (or, when the
+   parent Task call was seen, at that call). First lines written:
+   `Status Running`, then `UserMessage { text: <the Task's prompt> }` when
+   the prompt is known -- it genuinely is the subagent's first user turn.
+2. Every child line is translated by that subagent's own `Translator`
+   (one per subagent: tool ids are unique but streaming deltas are by
+   content-block index, and parallel subagents interleave).
+3. **The parent's `tool_result` never finishes a subagent.** The Task tool
+   runs in the background by default: the `tool_result` -- "Async agent
+   launched..." -- arrives the moment it *starts*, while the subagent goes
+   on working for however long its own turn takes, sometimes minutes. What
+   ends it is its own turn ending: the raw API's `message_delta` on its
+   stream carrying `stop_reason: "end_turn"` (a `stop_reason` of `tool_use`
+   is the model about to call one, not an end), or a `result` line for its
+   own turn if a future CLI version ever sends one. Either maps to
+   `Status Exited`; the subagent's vocabulary has no `Idle`, so the
+   equivalent event `dispatch` produces for an ordinary session is dropped
+   rather than written. A shipped version of this finished on the
+   `tool_result` instead, which read a running background agent as
+   "finished" with its transcript truncated at the moment it launched.
+4. **A child line for a subagent that already finished reopens it**
+   (`Status Running`) rather than being dropped: a background Task can be
+   sent another message long after its first turn ended, and that is
+   exactly what a further line for it means. Same transcript, same child
+   `Translator`, just picking back up.
+5. When the parent session's process exits (`Status Exited` on the
+   session), every subagent still `Running` gets `Status Exited` too: its
+   process was the parent's.
+
+A subagent that was mid-flight when the backend restarted keeps working:
+the registry reopens the existing transcript on the next child line, and
+the file continues its sequence -- the same reopening #4 describes, whether
+what closed it was a restart or its own `end_turn`. If its turn ended while
+the backend was down nothing recorded that until the next line arrives, so
+its last status stays `Running`, which the list reports as **unknown**
+rather than as running (see the wire shape) until then.
+
+Title: the Task call's `description` input, then ` (<subagent_type>)` when
+one is given; falling back to the tool's name when the child arrives before
+(or without) the parent call being seen.
+
+## Server layout
+
+- `session/subagent.rs` -- the registry: `Subagents` (per session, in
+  `Shared`), `Subagent` (its `Transcript` behind a mutex plus a
+  `broadcast::Sender<SeqEvent>`), `record(id, event)`, `start(id, title,
+  prompt)`, `finish(id)`, `reopen(id)`, `finish_all()`, `list()` from disk. Drivers get an
+  `Arc<Subagents>` beside their `EventSink`; llama ignores it.
+- `session/claude/translate.rs` -- routes child lines by parent id, holds
+  one child `Translator` per subagent, remembers pending Task calls'
+  description/prompt/subagent_type.
+- `session/echo.rs` -- `/subagent [n]`: the test rig. Starts *n* (default 1)
+  subagents at once, each named "helper k". Each writes the prompt as its
+  user message, streams a few words of text, runs one `Bash` tool call, then
+  finishes about three seconds after starting, and the parent's Task calls
+  end when their subagent does. Three seconds so the running state can be
+  seen on the phone.
+- `routes.rs` -- three routes, in the doc table.
+
+## Wire shape
+
+```
+GET /sessions/{id}                       SessionInfo gains `subagents: N` (count, 0 when none)
+GET /sessions                            same field on each row
+GET /sessions/{id}/subagents             [{id, title, status, created, lastActivity}], oldest first
+GET /sessions/{id}/subagents/{sub}/transcript   exactly the session transcript's query and answer
+GET /sessions/{id}/subagents/{sub}/events?after=N  exactly the session events stream
+```
+
+`status` is the transcript's last `Status` event, serialised like a session's
+(`running`, `exited`), except that a subagent whose session is not itself
+running cannot be running: the list answers `unknown` for that one. The
+phone words these as *running*, *finished* and *unknown* on the subcard.
+
+The count on `SessionInfo` is a directory listing, so the list stays cheap.
+The per-subagent status is only read when the list route is asked for.
+
+## Phone
+
+- `SessionSummary.subagents: Int`. A card with a non-zero count ends in an
+  expander row -- a full-width `Chevron(Pointing.Down)` row that flips to
+  `Pointing.Up` -- collapsed by default. Expanding fetches
+  `/sessions/{id}/subagents` and draws one `OutlinedCard` per subagent,
+  indented inside the session card, the way dev-updater draws a project's
+  components: title, then the status word and a relative time. The
+  expansion state is per session id and survives a refresh of the list.
+- Tapping a subcard opens `Screen.Subagent`, which is `SessionScreen` in
+  **read-only** form: the same transcript, paging, cache, selection,
+  images and status row, with the composer, the process button, the model
+  picker, the files button, the settings cog and the usage bar left out.
+  The header shows the subagent's title with the session's title beneath
+  it. Back returns to the list.
+- Addressing: `fetchTranscript`, `EventStream`, `TranscriptSource` and the
+  cache take a transcript address rather than a session id --
+  `sessions/{id}` or `sessions/{id}/subagents/{sub}` -- so the cache nests a
+  subagent's copy under its session's and the same code serves both.
@@ -102,6 +102,16 @@ android {
        targetSdk = 37
        versionCode = 1
        versionName = "1.0"
+        // Read by MainActivity to decide, at startup, whether this is the P0 benchmark build
+        // (docs/RUST.md's P0 box) rather than the app somebody enrolled. False everywhere except
+        // the `bench` build type below, which overrides it.
+        buildConfigField("boolean", "FIXTURE_MODE", "false")
+    }
+    buildFeatures {
+        // Only for FIXTURE_MODE above; nothing else here reaches for generated BuildConfig fields.
+        buildConfig = true
+        // Only for the bench build type's resValue("string", "app_name", ...) below.
+        resValues = true
    }
    packaging {
        resources { excludes += "/META-INF/{AL2.0,LGPL2.1}" }
@@ -131,6 +141,35 @@ android {
            isMinifyEnabled = false
            if (keystore != null) signingConfig = signingConfigs.getByName("release")
        }
+        // P0's benchmark build (docs/RUST.md, docs/DECISIONS.md's 2026-09-05 entry): release
+        // optimisations so a frame time measured here means what release means everywhere else in
+        // this project, its own application id so it installs beside a real enrollment rather than
+        // replacing it, and FIXTURE_MODE so MainActivity opens straight onto the fixture session
+        // instead of asking to be enrolled. Signed with the same key as release -- it never talks
+        // to a real backend, so there is no CA of its own to mismatch, and a second keystore would
+        // be one more secret to keep off this machine's shared mount for no benefit.
+        create("bench") {
+            initWith(getByName("release"))
+            // :link (wg-app-link) has no "bench" build type of its own -- it is a library shared
+            // with dev-updater and has no reason to know this project invented one -- so this says
+            // which of its build types to link against instead.
+            matchingFallbacks += listOf("release")
+            applicationIdSuffix = ".bench"
+            // "AI Sessions bench" everywhere the OS shows the app's name (launcher, recents,
+            // Settings): this resValue overrides res/values/strings.xml's app_name for this
+            // build type alone, and AndroidManifest.xml's android:label reads @string/app_name
+            // rather than a literal so a build type can override it without touching the
+            // manifest.
+            resValue("string", "app_name", "AI Sessions bench")
+            buildConfigField("boolean", "FIXTURE_MODE", "true")
+            if (keystore != null) signingConfig = signingConfigs.getByName("release")
+        }
+    }
+    sourceSets {
+        // The fixture both bench builds (this one and iris's) open with; see
+        // app/bench-fixture/README.md. Read directly from its own directory rather than copied
+        // into androidApp/src -- one file to keep in sync with the generator, not two.
+        getByName("bench").assets.directories.add("../bench-fixture/assets")
    }
    compileOptions {
        sourceCompatibility = JavaVersion.VERSION_21
@@ -25,7 +25,7 @@
         the fix is a judgement about how this app should look. Drop this
         suppression when a real icon lands. -->
    <application
-        android:label="AI Sessions"
+        android:label="@string/app_name"
        android:allowBackup="true"
        android:theme="@android:style/Theme.Material.Light.NoActionBar"
        tools:ignore="MissingApplicationIcon">
@@ -142,6 +142,20 @@ data class SessionSummary(
    val keepsOwnTranscript: Boolean,
    /** How much the session asks before acting; null when it was never set. */
    val permissionMode: String?,
+    /**
+     * How hard the model thinks, or null for the CLI's own default.
+     *
+     * Null is a level somebody can choose, not only one to start in -- see [EFFORT_LEVELS]. It is
+     * reported rather than assumed for the same reason [permissionMode] is.
+     */
+    val effort: String?,
+    /**
+     * Whether a thinking level does anything here -- a Claude CLI session, not a llama or echo one.
+     *
+     * Asked of the server rather than worked out from the provider's name, because this is a
+     * property of the driver's *kind* and the phone only has the name.
+     */
+    val takesEffort: Boolean,
    /**
     * Whether this continues a session the machine already had, which changes what deleting means.
     */
@@ -153,6 +167,24 @@ data class SessionSummary(
     * itself from a default is one you can turn off while believing you are reading it.
     */
    val notify: Boolean,
+    /**
+     * Whether this session sends itself a message once its account's usage limit lifts, and what
+     * that message says.
+     *
+     * The message is what the server would actually send, with its own default already filled in,
+     * so the field shows the words rather than an empty box standing for them.
+     */
+    val autoResume: Boolean,
+    val autoResumeMessage: String,
+    /**
+     * When the server next intends to check whether the limit has lifted, in epoch seconds, or null
+     * when nothing is waiting.
+     *
+     * A time to *ask*, not a time to resume: the server checks the meter at that moment and waits
+     * again if the limit is still on. Worded that way wherever it is shown, because a promise this
+     * app cannot keep is worse than no time at all.
+     */
+    val resumeAt: Double?,
    /**
     * The directory the session works in, or null where it was never given one.
     *
@@ -178,8 +210,27 @@ data class SessionSummary(
     * server because that is where a provider's kind is known.
     */
    val maxImageEdge: Int?,
+    /**
+     * Which of `GET /usage`'s snapshots is about this session, and null where nothing meters it.
+     *
+     * The rate-limit bar answers a question about an *account*, and what decides which account --
+     * if any -- is the provider this session runs, not the machine it runs on. Pairing by machine
+     * alone drew the Claude CLI's five-hour window under every echo session on a machine that also
+     * has the CLI: a quota that session cannot spend and could never run down. Decided by the
+     * server for the same reason [maxImageEdge] is -- it is a fact about the provider's kind, and
+     * this app has only its name.
+     */
+    val usageProvider: String?,
    val status: String,
    val lastActivity: Double,
+    /**
+     * How many subagents this session has, however their own status now reads.
+     *
+     * A directory listing on the server rather than a status read per subagent, so the list stays
+     * cheap; the per-subagent state is only fetched when the card is expanded. Zero on a server
+     * that predates subagents, so this app still opens against one.
+     */
+    val subagents: Int,
 )

 private fun parseSession(session: JSONObject) =
@@ -192,14 +243,24 @@ private fun parseSession(session: JSONObject) =
        title = session.getString("title"),
        model = session.optString("model").ifEmpty { null },
        permissionMode = session.optString("permissionMode").ifEmpty { null },
+        effort = session.optString("effort").ifEmpty { null },
+        takesEffort = session.optBoolean("takesEffort", false),
        imported = session.optBoolean("imported", false),
        notify = session.optBoolean("notify", true),
+        autoResume = session.optBoolean("autoResume", false),
+        // The server sends its own default rather than nothing, so an empty answer means an older
+        // server -- and this app's word for it is the same word.
+        autoResumeMessage =
+            session.optString("autoResumeMessage").ifEmpty { DEFAULT_RESUME_MESSAGE },
+        resumeAt = if (session.has("resumeAt")) session.getDouble("resumeAt") else null,
        cwd = session.optString("cwd").ifEmpty { null },
        contextTokens =
            if (session.has("contextTokens")) session.getLong("contextTokens") else null,
        maxImageEdge = session.optInt("maxImageEdge", 0).takeIf { it > 0 },
+        usageProvider = session.optString("usageProvider").ifEmpty { null },
        status = session.getString("status"),
        lastActivity = session.getDouble("lastActivity"),
+        subagents = session.optInt("subagents", 0),
    )

 fun fetchSessions(settings: ServerSettings): List<SessionSummary> =
@@ -215,6 +276,35 @@ fun fetchSessions(settings: ServerSettings): List<SessionSummary> =
 fun fetchSession(settings: ServerSettings, sessionId: String): SessionSummary =
    requestFromServer(settings, "/sessions/$sessionId") { parseSession(it.jsonObject()) }

+/**
+ * One row of `GET /sessions/{id}/subagents`, oldest first.
+ *
+ * A subagent is a second transcript owned by a session -- no process, no controls of its own -- so
+ * this carries only what a card needs to draw and to open it; see SUBAGENTS.md. [status] is
+ * "running", "exited" or "unknown": a subagent whose session is not itself running cannot be
+ * running, and the list says so rather than reporting a state that cannot hold.
+ */
+data class SubagentSummary(
+    val id: String,
+    val title: String,
+    val status: String,
+    val created: Double,
+    val lastActivity: Double,
+)
+
+fun fetchSubagents(settings: ServerSettings, sessionId: String): List<SubagentSummary> =
+    requestFromServer(settings, "/sessions/$sessionId/subagents") {
+        it.jsonObjects { row ->
+            SubagentSummary(
+                id = row.getString("id"),
+                title = row.getString("title"),
+                status = row.getString("status"),
+                created = row.getDouble("created"),
+                lastActivity = row.getDouble("lastActivity"),
+            )
+        }
+    }
+
 // What the server offers, so the spawn screen has no hardcoded lists: a setup added to the server's
 // config.ron appears here with no app rebuild.
 //
@@ -375,6 +465,12 @@ data class SshDetails(
     * Where files attached from here land on that machine; null for the session's own directory.
     */
    val attachmentsDir: String? = null,
+    /**
+     * Where that machine keeps its GGUF models; null for the same place the backend keeps its own
+     * (`~/.local/share/ai-app/models`, read on that machine). A llama.cpp session serves the file
+     * from the machine it runs on, so this is where its models are looked for and listed.
+     */
+    val modelsDir: String? = null,
 )

 private fun SshDetails.toJson() =
@@ -382,6 +478,7 @@ private fun SshDetails.toJson() =
        if (port != null) put("port", port)
        if (!identityFile.isNullOrBlank()) put("identityFile", identityFile)
        if (!attachmentsDir.isNullOrBlank()) put("attachmentsDir", attachmentsDir)
+        if (!modelsDir.isNullOrBlank()) put("modelsDir", modelsDir)
    }

 /** What a machine turns out to have, without saving anything. */
@@ -450,6 +547,8 @@ fun spawnSession(
    model: String? = null,
    cwd: String? = null,
    permissionMode: String? = null,
+    /** Null for whatever the server's default is; see [fetchDefaultEffort]. */
+    effort: String? = null,
    params: Map<String, String> = emptyMap(),
    /** Continue this Claude Code session instead of starting an empty one. */
    import: String? = null,
@@ -467,6 +566,7 @@ fun spawnSession(
                    if (!model.isNullOrBlank()) put("model", model)
                    if (!cwd.isNullOrBlank()) put("cwd", cwd)
                    if (!permissionMode.isNullOrBlank()) put("permissionMode", permissionMode)
+                    if (!effort.isNullOrBlank()) put("effort", effort)
                    if (!import.isNullOrBlank()) put("import", import)
                    if (params.isNotEmpty()) {
                        put("params", JSONObject(params.toMap<String, Any>()))
@@ -906,7 +1006,7 @@ fun startImport(
 */
 fun fetchTranscript(
    settings: ServerSettings,
-    sessionId: String,
+    address: TranscriptAddress,
    before: Long? = null,
    limit: Int = 80,
    // Count [limit] in rows, not events, joining a reply's streamed deltas into one -- so a page of
@@ -925,7 +1025,7 @@ fun fetchTranscript(
        if (coalesce) append("&coalesce=true")
        if (after != null) append("&after=").append(after)
    }
-    return requestFromServer(settings, "/sessions/$sessionId/transcript$query") { connection ->
+    return requestFromServer(settings, "/${address.urlPath}/transcript$query") { connection ->
        val body = JSONArray(connection.inputStream.bufferedReader().readText())
        // The text as well as the event: the transcript cache stores the one and the fold needs the
        // other, and they have to be the same line.
@@ -973,6 +1073,56 @@ fun setSessionModel(settings: ServerSettings, sessionId: String, model: String)
 */
 val PERMISSION_MODES = listOf("manual", "acceptEdits", "auto", "bypassPermissions", "plan")

+/**
+ * What a new session's thinking level is when nothing chose one, or null for the CLI's own.
+ *
+ * Held by the server rather than by this phone, because a second device would otherwise spawn
+ * sessions at a level the first one's owner never picked.
+ */
+fun fetchDefaultEffort(settings: ServerSettings): String? =
+    requestFromServer(settings, "/defaults") {
+        it.jsonObject().optString("effort").ifEmpty { null }
+    }
+
+/** Sets what new sessions start at. Nothing already running changes. */
+fun setDefaultEffort(settings: ServerSettings, level: String?) {
+    requestFromServer(
+        settings,
+        "/defaults",
+        method = "POST",
+        jsonBody = JSONObject().put("effort", level ?: JSONObject.NULL).toString(),
+    ) {}
+}
+
+/**
+ * How hard the model thinks, as `claude --effort` takes them, cheapest first.
+ *
+ * Not offered alongside the model and the permission mode on the session's own bar, because it does
+ * not behave like them: the CLI has a control request for those two and none for this (checked
+ * against 2.1.258), so a level is settled when the process is launched. Changing it therefore stops
+ * the process, which is what the working directory beside it in this dialog does, and why it is
+ * here rather than on a bar whose other controls take effect mid-turn.
+ */
+val EFFORT_LEVELS = listOf("low", "medium", "high", "xhigh", "max")
+
+/** What the picker shows, and sends as null, for a session that has chosen no level. */
+const val DEFAULT_EFFORT = "default"
+
+/**
+ * Records how hard a session thinks and **stops its process**, since the level is read when the
+ * process is launched. The next message, or Start, runs one that has it.
+ *
+ * [level] is null for the CLI's own default.
+ */
+fun setSessionEffort(settings: ServerSettings, sessionId: String, level: String?) {
+    requestFromServer(
+        settings,
+        "/sessions/$sessionId/effort",
+        method = "POST",
+        jsonBody = JSONObject().put("effort", level ?: JSONObject.NULL).toString(),
+    ) {}
+}
+
 /** Switches how much a running session asks before acting, also in place. */
 fun setSessionPermissionMode(settings: ServerSettings, sessionId: String, mode: String) {
    requestFromServer(
@@ -984,6 +1134,36 @@ fun setSessionPermissionMode(settings: ServerSettings, sessionId: String, mode:
 }

 /** Turns this session's notifications on or off. Stored on the backend -- see `SessionConfig`. */
+/**
+ * What an auto-resume says when nothing else was typed. Mirrors the server's own default, so a
+ * cleared field shows the word that would actually be sent instead of going blank.
+ */
+const val DEFAULT_RESUME_MESSAGE = "continue"
+
+/**
+ * Turns auto-resume on or off and sets what it would say, in one request because they are one
+ * decision -- see the server's `/sessions/{id}/auto-resume`.
+ */
+fun setSessionAutoResume(
+    settings: ServerSettings,
+    sessionId: String,
+    autoResume: Boolean,
+    message: String?,
+) {
+    requestFromServer(
+        settings,
+        "/sessions/$sessionId/auto-resume",
+        method = "POST",
+        jsonBody =
+            JSONObject()
+                .put("autoResume", autoResume)
+                // Empty means the server's default rather than a session poked with nothing to
+                // read, which is the same rule the server applies to the field.
+                .put("message", message?.trim()?.ifEmpty { null } ?: JSONObject.NULL)
+                .toString(),
+    ) {}
+}
+
 fun setSessionNotify(settings: ServerSettings, sessionId: String, notify: Boolean) {
    requestFromServer(
        settings,
@@ -1074,6 +1254,26 @@ private fun parseDownload(o: JSONObject) =
        error = if (o.has("error")) o.getString("error") else null,
    )

+/**
+ * The models on one machine, which is the list a llama.cpp session there can choose from.
+ *
+ * Not [fetchModels], which is what the *backend* has downloaded. A session serves its model from
+ * the machine it runs on, so for a machine reached over ssh those are two different lists -- and
+ * offering the backend's would name files that are not there, turning a choice that cannot work
+ * into a session that fails when it tries to load one.
+ */
+fun fetchSetupModels(settings: ServerSettings, setupId: String): List<LocalModel> =
+    requestFromServer(settings, "/setups/${setupId.urlEncoded()}/models") { connection ->
+        JSONArray(connection.inputStream.bufferedReader().readText()).mapObjects { m ->
+            LocalModel(
+                key = m.getString("key"),
+                repo = m.getString("repo"),
+                file = m.getString("file"),
+                bytes = m.getLong("bytes"),
+            )
+        }
+    }
+
 fun fetchModels(settings: ServerSettings): Models =
    requestFromServer(settings, "/models") { connection ->
        val body = JSONObject(connection.inputStream.bufferedReader().readText())
@@ -1,7 +1,9 @@
 package com.example.aiapp

 import androidx.activity.compose.BackHandler
+import androidx.compose.foundation.background
 import androidx.compose.foundation.layout.Box
+import androidx.compose.foundation.layout.fillMaxSize
 import androidx.compose.foundation.layout.imePadding
 import androidx.compose.foundation.layout.padding
 import androidx.compose.material3.AlertDialog
@@ -34,7 +36,23 @@ import kotlinx.coroutines.withContext
 * session, spawning one, and settings.
 */
 private sealed class Screen {
-    data object Main : Screen()
+    /**
+     * The session list, with a subagent's own transcript over it when [subagent] is set.
+     *
+     * A layer on this screen rather than a screen of its own, for the same reason [Session.files]
+     * is: [SessionListScreen] owns which cards are expanded and what each expansion fetched, kept
+     * in `remember`, and a subagent is opened from a card's expander. As a sibling `Screen` it was
+     * disposed and recreated on every return, which lost that state -- an expanded card collapsed
+     * itself the moment its own subagent's view was closed.
+     */
+    data class Main(val subagent: SubagentTarget? = null) : Screen()
+
+    /**
+     * One subagent's own transcript, read-only. See [SessionScreen]'s `subagent` parameter and
+     * SUBAGENTS.md's "Phone". Closing it returns to [Main] under it, not to [Session]: a subagent
+     * is opened from the session list's card rather than from inside the session it belongs to.
+     */
+    data class SubagentTarget(val summary: SessionSummary, val subagent: SubagentSummary)

    /**
     * One session, with the file explorer over it when [files] is set.
@@ -81,7 +99,7 @@ fun AppRoot(
    val context = LocalContext.current
    val scope = rememberCoroutineScope()
    var settings by remember(settingsVersion) { mutableStateOf(loadServerSettings(context)) }
-    var screen by remember { mutableStateOf<Screen>(Screen.Main) }
+    var screen by remember { mutableStateOf<Screen>(Screen.Main()) }
    // A notification tap this could not follow, and why. Null both before one is asked for and
    // after one succeeds, since success is a screen rather than a message.
    var failedOpen by remember { mutableStateOf<FailedOpen?>(null) }
@@ -96,7 +114,7 @@ fun AppRoot(
            share = shareRequest
            // A session already open takes it. Otherwise the list is where the choice is made,
            // whatever screen was showing: Spawn and Settings have nowhere to put a file.
-            if (screen !is Screen.Session) screen = Screen.Main
+            if (screen !is Screen.Session) screen = Screen.Main()
        }
    }

@@ -123,7 +141,7 @@ fun AppRoot(
                existing = null,
                onSaved = { saved ->
                    settings = saved
-                    screen = Screen.Main
+                    screen = Screen.Main()
                },
                onBack = null,
            )
@@ -136,7 +154,7 @@ fun AppRoot(
    // shows, so it always refetches.
    val goToMain = {
        reloadToken++
-        screen = Screen.Main
+        screen = Screen.Main()
    }
    if (screen !is Screen.Main) {
        BackHandler(onBack = goToMain)
@@ -185,6 +203,9 @@ fun AppRoot(
                    reloadToken = reloadToken,
                    share = share,
                    onOpen = { screen = Screen.Session(it) },
+                    onOpenSubagent = { summary, subagent ->
+                        screen = here.copy(subagent = Screen.SubagentTarget(summary, subagent))
+                    },
                    onSpawn = { screen = Screen.Spawn },
                    onImported = { imported ->
                        reloadToken++
@@ -192,6 +213,27 @@ fun AppRoot(
                    },
                    onSettings = { screen = Screen.Settings },
                )
+                // Its own back handler is registered after MainScreen's, so it is the one the
+                // platform asks first while a subagent is open -- the same rule the files
+                // explorer's handler follows over its session, below.
+                here.subagent?.let { target ->
+                    BackHandler { screen = here.copy(subagent = null) }
+                    // Its own opaque background: this screen was always the sole content under
+                    // the theme's own Surface before, so it never had to paint one -- stacked over
+                    // the list here, the space between its own cards let the list underneath show
+                    // through without this. The same fix FilesScreen needed over its session.
+                    Box(Modifier.fillMaxSize().background(MaterialTheme.colorScheme.background)) {
+                        key(target.summary.id, target.subagent.id) {
+                            SessionScreen(
+                                settings = current,
+                                summary = target.summary,
+                                onBack = { screen = here.copy(subagent = null) },
+                                onFiles = {},
+                                subagent = target.subagent,
+                            )
+                        }
+                    }
+                }
            }
        is Screen.Session ->
            // Keyed on the id, because a different session is a different screen rather than this
@@ -0,0 +1,108 @@
+package com.example.aiapp
+
+import android.content.Context
+import java.util.concurrent.CopyOnWriteArrayList
+
+/**
+ * P0's benchmark gate (see docs/RUST.md and docs/DECISIONS.md's 2026-09-05 entry): an in-process
+ * fake of the backend, so the `bench` build type can drive a real session screen -- the real
+ * [TranscriptSource], the real fold, the real paging -- with no server and no network permission.
+ *
+ * Only ever installed when [BuildConfig.FIXTURE_MODE] is true (see [MainActivity]); everything else
+ * in this build compiles it in but never calls it, since Kotlin has no per-build-type source set
+ * that both [MainActivity] (which every variant compiles) and this can share without one.
+ *
+ * The design: [requestFromServer] and [Sse] talk to `https://$FIXTURE_HOST:$FIXTURE_PORT` through
+ * ordinary `java.net.URL`, exactly as they would talk to a real server. A
+ * [java.net.URLStreamHandlerFactory] registered once for the whole process intercepts every
+ * `https://` connection to that host and answers from this object's in-memory event log instead of
+ * opening a socket -- see BenchNetwork.kt. Everything above that (TranscriptSource, SessionScreen,
+ * the fold, uniqueItems) never learns the difference.
+ */
+object BenchFixture {
+    const val FIXTURE_HOST = "bench.fixture.invalid"
+    const val FIXTURE_PORT = 1
+
+    /** How many of the fixture's events are the opening backlog; see bench-fixture/README.md. */
+    private const val BACKLOG_COUNT = 3200
+
+    val settings = ServerSettings(FIXTURE_HOST, FIXTURE_PORT, "bench")
+
+    /** The session id every bench run opens; nothing else in this build ever mints one. */
+    const val SESSION_ID = "bench-fixture-session"
+
+    /**
+     * The whole transcript, seq order, growing as [pushLive] is called during the streaming phase.
+     * Read by both the REST page handler and the SSE handler, so a page requested mid- stream and a
+     * live frame agree on what has "already happened" -- the same thing a real server's own
+     * transcript file guarantees.
+     */
+    private val log = CopyOnWriteArrayList<Pair<String, SeqEvent>>()
+
+    /** The events not yet appended to [log] -- the streaming phase's own source. */
+    private var streamTail: List<Pair<String, SeqEvent>> = emptyList()
+
+    private val images = mutableMapOf<String, ByteArray>()
+
+    @Volatile private var loaded = false
+
+    /**
+     * Parses the bundled fixture once. Safe to call more than once; only the first does anything.
+     */
+    @Synchronized
+    fun ensureLoaded(context: Context) {
+        if (loaded) return
+        val lines =
+            context.assets.open("transcript.jsonl").bufferedReader().readLines().filter {
+                it.isNotBlank()
+            }
+        val parsed = lines.map { it to parseSeqEvent(it) }
+        log.addAll(parsed.take(BACKLOG_COUNT))
+        streamTail = parsed.drop(BACKLOG_COUNT)
+        for (name in listOf("bench1.png", "bench2.png")) {
+            images[name] = context.assets.open(name).readBytes()
+        }
+        loaded = true
+    }
+
+    /** The events the streaming phase has left to send. */
+    fun remainingStreamEvents(): Int = streamTail.size
+
+    /** Sends the next fixture event onto the live log, as a real SSE frame would arrive. */
+    fun pushNextLiveEvent(): Boolean {
+        val next = streamTail.firstOrNull() ?: return false
+        streamTail = streamTail.drop(1)
+        log.add(next)
+        return true
+    }
+
+    /** Undoes [pushNextLiveEvent] and reloads the opening backlog, for running the bench twice. */
+    @Synchronized
+    fun resetToBacklog(context: Context) {
+        loaded = false
+        log.clear()
+        ensureLoaded(context)
+    }
+
+    fun fileBytes(name: String): ByteArray? = images[name]
+
+    /**
+     * Raw JSON lines with seq > [after], in order -- what an `/events?after=` connection replays.
+     */
+    fun linesAfter(after: Long): List<String> =
+        log.filter { it.second.seq > after }.map { it.first }
+
+    /**
+     * One REST page: [fetchTranscript]'s `before`/`limit`/`after`, against the growing log. Ignores
+     * `coalesce` -- the fixture's own deltas are already split the way a real reply streams, and
+     * what the benchmark exercises is the fold and the paging, not the server's row-joining, which
+     * client-core's own port tracks separately (CLIENT_CORE.md).
+     */
+    fun page(before: Long?, limit: Int, after: Long?): List<String> {
+        val upper = before ?: (log.lastOrNull()?.second?.seq?.plus(1) ?: 1L)
+        val candidates = log.filter {
+            it.second.seq < upper && (after == null || it.second.seq > after)
+        }
+        return candidates.takeLast(limit).map { it.first }
+    }
+}
@@ -0,0 +1,181 @@
+package com.example.aiapp
+
+import java.io.ByteArrayInputStream
+import java.io.IOException
+import java.io.InputStream
+import java.io.PipedInputStream
+import java.io.PipedOutputStream
+import java.net.HttpURLConnection
+import java.net.URL
+import java.net.URLStreamHandler
+import java.net.URLStreamHandlerFactory
+import java.security.Principal
+import java.security.cert.Certificate
+import javax.net.ssl.HttpsURLConnection
+import javax.net.ssl.SSLPeerUnverifiedException
+import org.json.JSONArray
+
+/**
+ * Installs the process-wide interception [BenchFixture] needs. Idempotent and safe to call more
+ * than once; the JDK only allows [URL.setURLStreamHandlerFactory] to be called successfully once
+ * per process, and a second real call throws -- so this guards it rather than relying on every
+ * caller to remember.
+ *
+ * Scoped to [BenchFixture.FIXTURE_HOST]: any other `https://` URL falls through to the platform's
+ * ordinary handler, so this only ever changes behaviour for the one host the bench build invents.
+ */
+@Synchronized
+fun installFixtureNetworkOnce() {
+    if (installed) return
+    installed = true
+    URL.setURLStreamHandlerFactory(
+        URLStreamHandlerFactory { protocol ->
+            if (protocol != "https") null
+            else
+                object : URLStreamHandler() {
+                    override fun openConnection(url: URL): HttpURLConnection =
+                        if (url.host == BenchFixture.FIXTURE_HOST) FixtureConnection(url)
+                        else
+                        // The bench build makes no other https call -- this factory is
+                        // installed only in FIXTURE_MODE (MainActivity) -- so there is
+                        // deliberately no delegate to a platform handler here: once a
+                        // URLStreamHandlerFactory is installed there is no supported way to
+                        // ask the JDK for its own default handler back, and re-entering this
+                        // same factory for the fallback would recurse forever rather than
+                        // reach one.
+                        throw java.io.IOException(
+                                "bench build's fixture network has no route to https host " +
+                                    "${url.host} -- only ${BenchFixture.FIXTURE_HOST} is served"
+                            )
+                }
+        }
+    )
+}
+
+private var installed = false
+
+/**
+ * Answers one request against [BenchFixture] instead of opening a socket. Implements just enough of
+ * [HttpsURLConnection] for [requestFromServer] and [Sse] to work unmodified: both only call
+ * `connect`/`disconnect`, set a handful of request properties they never need answered, and read
+ * `responseCode` and `inputStream`.
+ */
+private class FixtureConnection(url: URL) : HttpsURLConnection(url) {
+    private var input: InputStream? = null
+    private var writer: Thread? = null
+
+    override fun connect() {
+        if (input != null) return
+        input = route(url.path, url.query)
+    }
+
+    override fun disconnect() {
+        writer?.interrupt()
+        try {
+            input?.close()
+        } catch (_: IOException) {}
+    }
+
+    override fun usingProxy() = false
+
+    override fun getResponseCode(): Int {
+        connect()
+        return 200
+    }
+
+    override fun getInputStream(): InputStream {
+        connect()
+        return input!!
+    }
+
+    override fun getErrorStream(): InputStream? = null
+
+    // Nothing here reads any of these; implemented only because HttpsURLConnection declares them
+    // abstract. A fixture never negotiates real TLS, so each says exactly that rather than
+    // fabricating a plausible-looking certificate.
+    override fun getCipherSuite() = "none (bench fixture, no TLS)"
+
+    override fun getLocalCertificates(): Array<Certificate>? = null
+
+    override fun getServerCertificates(): Array<Certificate> =
+        throw SSLPeerUnverifiedException("bench fixture connection presents no certificate")
+
+    override fun getPeerPrincipal(): Principal =
+        throw SSLPeerUnverifiedException("bench fixture connection presents no certificate")
+
+    override fun getLocalPrincipal(): Principal? = null
+
+    /**
+     * [path] is `/sessions/{id}/...`; everything else this build's fixture is asked for is a bug.
+     */
+    private fun route(path: String, query: String?): InputStream {
+        val params =
+            (query ?: "")
+                .split("&")
+                .filter { it.contains('=') }
+                .associate {
+                    val (k, v) = it.split("=", limit = 2)
+                    k to java.net.URLDecoder.decode(v, "UTF-8")
+                }
+        return when {
+            path.endsWith("/transcript") -> {
+                val lines =
+                    BenchFixture.page(
+                        before = params["before"]?.toLongOrNull(),
+                        limit = params["limit"]?.toIntOrNull() ?: 80,
+                        after = params["after"]?.toLongOrNull(),
+                    )
+                val body = JSONArray(lines.map { org.json.JSONObject(it) })
+                ByteArrayInputStream(body.toString().toByteArray())
+            }
+            path.endsWith("/events") -> openEventsStream(params["after"]?.toLongOrNull() ?: 0L)
+            path.contains("/files/") -> {
+                val name = path.substringAfterLast("/files/")
+                val bytes =
+                    BenchFixture.fileBytes(name)
+                        ?: throw IOException("bench fixture has no file named $name")
+                ByteArrayInputStream(bytes)
+            }
+            else -> throw IOException("bench fixture has no route for $path")
+        }
+    }
+
+    /**
+     * A live SSE body: [BenchFixture.linesAfter] replayed immediately, then polled every 50ms for
+     * anything [BenchFixture.pushNextLiveEvent] has added since -- the same shape a real backend's
+     * backlog-then-follow gives [Sse], just polled instead of woken, which is a fixture's business
+     * rather than something worth a condition variable for.
+     */
+    private fun openEventsStream(after: Long): InputStream {
+        val pipeIn = PipedInputStream(1 shl 16)
+        val pipeOut = PipedOutputStream(pipeIn)
+        var sent = after
+        val thread = Thread {
+            try {
+                while (!Thread.currentThread().isInterrupted) {
+                    val fresh = BenchFixture.linesAfter(sent)
+                    for (line in fresh) {
+                        pipeOut.write("data: $line\n\n".toByteArray())
+                        pipeOut.flush()
+                        sent = org.json.JSONObject(line).getLong("seq")
+                    }
+                    Thread.sleep(50)
+                }
+            } catch (_: InterruptedException) {
+                // disconnect() -- the ordinary way this ends.
+            } catch (_: IOException) {
+                // The reader side (Sse) closed its end.
+            } finally {
+                try {
+                    pipeOut.close()
+                } catch (_: IOException) {}
+            }
+        }
+            .also {
+                it.isDaemon = true
+                it.start()
+            }
+        writer = thread
+        return pipeIn
+    }
+}
@@ -0,0 +1,143 @@
+package com.example.aiapp
+
+import android.content.Context
+import android.os.BatteryManager
+import android.os.Process
+import androidx.compose.animation.core.tween
+import androidx.compose.foundation.gestures.animateScrollBy
+import androidx.compose.foundation.lazy.LazyListState
+import java.io.File
+import kotlinx.coroutines.CoroutineScope
+import kotlinx.coroutines.delay
+import kotlinx.coroutines.isActive
+import kotlinx.coroutines.launch
+
+/**
+ * P0's scripted benchmark, run in-process instead of by a shell script: the phone has no usable
+ * system tracing (this-machine-android's skill) and no agent can drive it, so the same scroll loop
+ * and streaming phase `transcript-bench.sh`/`stream-bench.sh` drive over `ui-trace` are reproduced
+ * here against [LazyListState] and [BenchFixture] directly. Only reachable from the `bench` build
+ * (see [SessionSettingsDialog]'s `onRunBenchmark`), but compiled into every build for the reason
+ * [BenchFixture]'s doc comment gives.
+ */
+object BenchRun {
+    /** transcript-bench.sh's default: 6 cycles of 4 swipes each, 900px over 200ms, 500ms apart. */
+    private const val CYCLES = 6
+    private const val SWIPE_PX = 900f
+    private const val SWIPE_MS = 200
+    private const val SWIPE_PAUSE_MS = 500L
+
+    /** stream-bench.sh's shape: a real reply arrives as many small deltas, not one big write. */
+    private const val STREAM_EVENTS_PER_SEC = 20
+    private const val STREAM_SECONDS = 20
+
+    /**
+     * Scrolls, then streams, then returns the extra report lines P0 asked for (CPU time, peak RSS,
+     * battery current) -- [FrameStats] and [DebugStats] are reset first, exactly as
+     * `copyRenderReport` resets them, so the two accountings cover the same stretch of work.
+     */
+    suspend fun run(
+        context: Context,
+        scope: CoroutineScope,
+        listState: LazyListState,
+    ): List<String> {
+        FrameStats.reset()
+        DebugStats.reset()
+        val cpuStartMs = Process.getElapsedCpuTime()
+
+        val battery = BatterySampler(context)
+        // Launched in the caller's scope rather than a fresh coroutineScope{} here, which would
+        // suspend this function until the sampler job ended -- and it only ends when told to.
+        val samplerJob = scope.launch {
+            while (isActive) {
+                battery.sample()
+                delay(1000)
+            }
+        }
+
+        // The swipe loop: transcript-bench.sh's four swipes per cycle are two drags toward newer
+        // content and two back, so a cycle returns to where it started and the whole loop measures
+        // steady-state scrolling rather than travelling somewhere new each time.
+        repeat(CYCLES) {
+            repeat(2) {
+                listState.animateScrollBy(SWIPE_PX, tween(SWIPE_MS))
+                delay(SWIPE_PAUSE_MS)
+            }
+            repeat(2) {
+                listState.animateScrollBy(-SWIPE_PX, tween(SWIPE_MS))
+                delay(SWIPE_PAUSE_MS)
+            }
+        }
+
+        // Pinned to the newest end before streaming starts, the way stream-bench.sh's "Jump to
+        // latest" tap is -- a reply streamed into a list parked further back arrives off-screen and
+        // the report would show nothing happened.
+        listState.scrollToItem(0)
+
+        var sent = 0
+        val total = STREAM_EVENTS_PER_SEC * STREAM_SECONDS
+        while (sent < total && BenchFixture.remainingStreamEvents() > 0) {
+            BenchFixture.pushNextLiveEvent()
+            sent++
+            delay(1000L / STREAM_EVENTS_PER_SEC)
+        }
+        // Lets the last few deltas land and draw before the report is read.
+        delay(300)
+
+        samplerJob.cancel()
+        val cpuMs = Process.getElapsedCpuTime() - cpuStartMs
+        val rssLine = peakRssLine()
+        val batteryLine = battery.finish()
+
+        return listOf(
+            "  scroll: $CYCLES cycles (${CYCLES * 4} swipes), streamed $sent/$total fixture events",
+            "  process CPU time over this run: ${cpuMs}ms",
+            rssLine,
+            batteryLine,
+        )
+    }
+
+    /** VmHWM from /proc/self/status: the process's high-water mark, in kB, since it started. */
+    private fun peakRssLine(): String {
+        val kb =
+            try {
+                File("/proc/self/status")
+                    .readLines()
+                    .firstOrNull { it.startsWith("VmHWM:") }
+                    ?.trim()
+                    ?.removePrefix("VmHWM:")
+                    ?.trim()
+                    ?.removeSuffix("kB")
+                    ?.trim()
+                    ?.toLongOrNull()
+            } catch (_: Exception) {
+                null
+            }
+        return "  peak RSS: " +
+            (kb?.let { "${it}kB" } ?: "unavailable (/proc/self/status unreadable)")
+    }
+}
+
+/**
+ * Samples [BatteryManager.BATTERY_PROPERTY_CURRENT_NOW] (microamps) once a second for the length of
+ * a run. The property returns `Int.MIN_VALUE` on hardware that does not support it -- most
+ * emulators -- and that is reported as "unavailable" rather than folded into an average with the
+ * real samples, which would silently understate every number after it. See UI_RULES: never present
+ * an inferred value as a measured one.
+ */
+private class BatterySampler(context: Context) {
+    private val manager = context.getSystemService(BatteryManager::class.java)
+    private val samples = mutableListOf<Int>()
+
+    fun sample() {
+        val value = manager?.getIntProperty(BatteryManager.BATTERY_PROPERTY_CURRENT_NOW)
+        if (value != null && value != Int.MIN_VALUE) samples.add(value)
+    }
+
+    fun finish(): String {
+        if (samples.isEmpty()) return "  battery current: unavailable on this device"
+        val meanUa = samples.sum() / samples.size
+        return "  battery current: mean ${meanUa}µA over ${samples.size} samples" +
+            " (min ${samples.min()}, max ${samples.max()})"
+    }
+}
@@ -127,6 +127,13 @@ fun debugReport(
    frames: List<String>,
    accounting: List<String>,
    crash: String?,
+    /**
+     * P0's benchmark-only measurements (process CPU time, peak RSS, battery current) -- empty on
+     * every path but [BenchRun.runP0Benchmark], which is the only caller that has them. A section
+     * heading only appears when there is something to put under it, so an ordinary copy from the
+     * render-report button reads exactly as it did before this existed.
+     */
+    extra: List<String> = emptyList(),
 ): String = buildString {
    appendLine("ai-app render report")
    appendLine(device)
@@ -152,6 +159,11 @@ fun debugReport(
    appendLine("work since this was last copied:")
    val work = DebugStats.lines()
    if (work.isEmpty()) appendLine("  nothing recorded") else work.forEach { appendLine(it) }
+    if (extra.isNotEmpty()) {
+        appendLine()
+        appendLine("bench:")
+        extra.forEach { appendLine(it) }
+    }
 }

 /** Puts [text] on the clipboard under [label], which is what the system offers as its name. */
@@ -12,6 +12,10 @@ import androidx.compose.ui.Alignment
 import androidx.compose.ui.Modifier
 import androidx.compose.ui.graphics.Color
 import androidx.compose.ui.unit.dp
+import java.time.Instant
+import java.time.ZoneId
+import java.time.format.DateTimeFormatter
+import java.time.format.FormatStyle

 /**
 * A line across the transcript saying what left the session's context.
@@ -47,3 +51,40 @@ fun TranscriptDivider(text: String, color: Color, modifier: Modifier = Modifier)
 fun ClearedRow(modifier: Modifier = Modifier) {
    TranscriptDivider("Context cleared", clearedColor, modifier)
 }
+
+/**
+ * The mark running out of quota leaves.
+ *
+ * The same red the usage bar takes when a window is spent, because it is the same fact in a second
+ * place: colour by consequence, so "there is nothing left to spend" is learned once.
+ *
+ * A time rather than a countdown. The row is folded once and never re-measured, so a span would go
+ * stale on screen the moment it was drawn; and this is when the *account* said it would reset,
+ * which is not a promise about when the session picks back up. A limit the session was told no
+ * reset time for says nothing about one -- that state has its own words rather than a plausible
+ * number.
+ */
+@Composable
+fun LimitRow(item: TranscriptItem.LimitNote, modifier: Modifier = Modifier) {
+    TranscriptDivider(limitSummary(item.resetsAt, ZoneId.systemDefault()), overLimitColor, modifier)
+}
+
+/**
+ * What the row says. Split out so the wording is testable without a screen, since the two states it
+ * has to keep apart -- a reset time that arrived and one that never did -- are exactly the pair
+ * that reads the same when it goes wrong.
+ *
+ * [zone] is a parameter rather than read here so a test says the same thing wherever it runs.
+ */
+fun limitSummary(resetsAt: Double?, zone: ZoneId): String {
+    val at = resetsAt?.let {
+        try {
+            DateTimeFormatter.ofLocalizedTime(FormatStyle.SHORT)
+                .withZone(zone)
+                .format(Instant.ofEpochSecond(it.toLong()))
+        } catch (_: Exception) {
+            null
+        }
+    }
+    return if (at == null) "Usage limit reached" else "Usage limit reached • resets $at"
+}
@@ -14,7 +14,7 @@ private const val RESET_EVENT = "reset"
 * mean. [close] from any thread ends it, and the caller owns reconnecting -- with the last seq it
 * saw as the new cursor.
 */
-class EventStream(settings: ServerSettings, private val sessionId: String) {
+class EventStream(settings: ServerSettings, private val address: TranscriptAddress) {
    private val stream = Sse(settings)

    fun close() = stream.close()
@@ -35,7 +35,7 @@ class EventStream(settings: ServerSettings, private val sessionId: String) {
        // one and the screen folds the other, and they have to be the same line.
        onEvent: (raw: String, event: SeqEvent) -> Unit,
    ) {
-        stream.run("/sessions/$sessionId/events?after=$after", onOpen) { name, data ->
+        stream.run("/${address.urlPath}/events?after=$after", onOpen) { name, data ->
            // A named frame carries no payload and a data frame has no name.
            if (name == RESET_EVENT) onReset()
            else if (data.isNotEmpty()) onEvent(data, parseSeqEvent(data))
@@ -157,6 +157,18 @@ sealed class SessionEvent {
     */
    data object Cleared : SessionEvent()

+    /**
+     * The session stopped because its account's usage limit was reached.
+     *
+     * Its own event rather than an [Error] carrying the CLI's sentence, because it is a state
+     * rather than something that went wrong -- and because the raw sentence is `Claude AI usage
+     * limit reached|1788546972`, which is not readable by the person it is shown to.
+     *
+     * [resetsAt] is epoch seconds and null where the session was told nothing. Only the server acts
+     * on it; what this draws it as is a time, not a countdown, because nothing here re-measures it.
+     */
+    data class LimitReached(val resetsAt: Double?) : SessionEvent()
+
    data class Error(val message: String) : SessionEvent()

    /**
@@ -261,6 +273,10 @@ fun parseSeqEvent(json: String): SeqEvent {
                    trigger = body.optString("trigger").ifEmpty { null },
                )
            "cleared" -> SessionEvent.Cleared
+            "limitReached" ->
+                SessionEvent.LimitReached(
+                    if (body.has("resetsAt")) body.getDouble("resetsAt") else null
+                )
            "error" -> SessionEvent.Error(body.getString("message"))
            else -> SessionEvent.Unknown(type)
        }
@@ -66,6 +66,35 @@ class MainActivity : ComponentActivity() {
        // Transparent status bar on every version; the Surface below paints through underneath it
        // and content insets itself. Same reasoning as dev-updater's MainActivity.
        enableEdgeToEdge()
+
+        // The `bench` build's entire purpose (P0, docs/RUST.md): open straight onto the session
+        // screen against BenchFixture's in-process fake backend, with no enrollment, no network
+        // permission, and no notification prompt -- none of them mean anything with no server and
+        // no real device to notify. See BenchFixture.kt and BenchNetwork.kt for how a screen built
+        // to talk to a real backend is made to talk to this instead. Still needs the same
+        // status/navigation-bar padding the ordinary flow below applies: edge-to-edge is the
+        // platform's own default from Android 15 on this app's targetSdk, with or without the call
+        // above, so skipping the padding here put the header's own buttons under the status bar --
+        // there to look at, but not there for `ui-trace`'s tap-by-label to land on.
+        if (BuildConfig.FIXTURE_MODE) {
+            installFixtureNetworkOnce()
+            BenchFixture.ensureLoaded(this)
+            setContent {
+                MaterialTheme(colorScheme = AiAppColors) {
+                    Surface(modifier = Modifier.fillMaxSize()) {
+                        Box(Modifier.fillMaxSize().statusBarsPadding().navigationBarsPadding()) {
+                            SessionScreen(
+                                settings = BenchFixture.settings,
+                                summary = benchSessionSummary(),
+                                onBack = { finish() },
+                                onFiles = {},
+                            )
+                        }
+                    }
+                }
+            }
+            return
+        }
        // Dark status-bar icons only over a light background, decided from the scheme rather than
        // fixed. It was hardcoded to `true`, which was right against the default light surface and
        // became unreadable the moment the app wore Catppuccin Mocha.
@@ -147,6 +176,26 @@ class MainActivity : ComponentActivity() {
        }
    }

+    /** The one session the `bench` build ever shows -- BenchFixture's session id, nothing else. */
+    private fun benchSessionSummary() =
+        SessionSummary(
+            id = BenchFixture.SESSION_ID,
+            setup = "bench",
+            setupName = "bench",
+            provider = "bench",
+            title = "P0 benchmark",
+            model = null,
+            keepsOwnTranscript = false,
+            permissionMode = null,
+            imported = false,
+            notify = false,
+            cwd = null,
+            contextTokens = null,
+            maxImageEdge = null,
+            status = "idle",
+            lastActivity = 0.0,
+        )
+
    // launchMode="singleTop": an enrollment scan, or a notification tapped while the app is open,
    // lands here rather than in a second activity instance.
    override fun onNewIntent(intent: Intent) {
@@ -48,6 +48,8 @@ fun MainScreen(
    /** What another app shared in and no session has taken yet; see [ShareRequest]. */
    share: ShareRequest? = null,
    onOpen: (SessionSummary) -> Unit,
+    /** Opens one session's subagent, from the expander under its card. */
+    onOpenSubagent: (SessionSummary, SubagentSummary) -> Unit,
    onSpawn: () -> Unit,
    onImported: (SessionSummary) -> Unit,
    onSettings: () -> Unit,
@@ -139,6 +141,7 @@ fun MainScreen(
                    settings = settings,
                    reloadToken = token,
                    onOpen = onOpen,
+                    onOpenSubagent = onOpenSubagent,
                    onSpawn = onSpawn,
                )
            MainTab.Import ->
@@ -1,7 +1,9 @@
 package com.example.aiapp

 import androidx.compose.foundation.ExperimentalFoundationApi
+import androidx.compose.foundation.clickable
 import androidx.compose.foundation.combinedClickable
+import androidx.compose.foundation.layout.Arrangement
 import androidx.compose.foundation.layout.Box
 import androidx.compose.foundation.layout.Column
 import androidx.compose.foundation.layout.Row
@@ -9,6 +11,7 @@ import androidx.compose.foundation.layout.Spacer
 import androidx.compose.foundation.layout.fillMaxSize
 import androidx.compose.foundation.layout.fillMaxWidth
 import androidx.compose.foundation.layout.height
+import androidx.compose.foundation.layout.heightIn
 import androidx.compose.foundation.layout.padding
 import androidx.compose.foundation.layout.width
 import androidx.compose.foundation.lazy.LazyColumn
@@ -17,6 +20,7 @@ import androidx.compose.material3.Card
 import androidx.compose.material3.CircularProgressIndicator
 import androidx.compose.material3.FloatingActionButton
 import androidx.compose.material3.MaterialTheme
+import androidx.compose.material3.OutlinedCard
 import androidx.compose.material3.Switch
 import androidx.compose.material3.Text
 import androidx.compose.material3.TextButton
@@ -30,6 +34,8 @@ import androidx.compose.runtime.setValue
 import androidx.compose.ui.Alignment
 import androidx.compose.ui.Modifier
 import androidx.compose.ui.platform.LocalContext
+import androidx.compose.ui.semantics.contentDescription
+import androidx.compose.ui.semantics.semantics
 import androidx.compose.ui.unit.dp
 import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.launch
@@ -47,12 +53,40 @@ fun SessionListScreen(
    settings: ServerSettings,
    reloadToken: Int,
    onOpen: (SessionSummary) -> Unit,
+    /** Opens one session's subagent, from the expander under its card. */
+    onOpenSubagent: (SessionSummary, SubagentSummary) -> Unit,
    onSpawn: () -> Unit,
 ) {
    val scope = rememberCoroutineScope()
    var listState by remember { mutableStateOf<LoadState<List<SessionSummary>>>(LoadState.Loading) }
    var confirmingDelete by remember { mutableStateOf<SessionSummary?>(null) }

+    // Which session cards are expanded to show their subagents, and what each expansion fetched.
+    // Ids rather than a flag on the row for the same reason `deleting` is: the rows are rebuilt
+    // from
+    // whatever the server last said, and this belongs to the reader's own choice, which survives a
+    // refresh.
+    var expandedSessions by remember { mutableStateOf(setOf<String>()) }
+    var subagentLoads by remember {
+        mutableStateOf(mapOf<String, LoadState<List<SubagentSummary>>>())
+    }
+
+    fun loadSubagents(sessionId: String) {
+        subagentLoads = subagentLoads + (sessionId to LoadState.Loading)
+        scope.launch {
+            subagentLoads =
+                subagentLoads +
+                    (sessionId to
+                        try {
+                            LoadState.Loaded(
+                                withContext(Dispatchers.IO) { fetchSubagents(settings, sessionId) }
+                            )
+                        } catch (e: ApiException) {
+                            LoadState.failed(e)
+                        })
+        }
+    }
+
    // Failures that belong to one session rather than to the list, keyed by its id and shown on its
    // own card. The two scopes are decided by whether the server answered: it answered and refused,
    // so this says nothing about the other rows.
@@ -84,6 +118,14 @@ fun SessionListScreen(
                    withContext(Dispatchers.IO) {
                        transcriptCache.retainOnly(loaded.value.map { it.id }.toSet())
                    }
+                    // A session gone from this answer cannot still be expanded, and an expanded one
+                    // that is still here asks again -- its subagents may have changed since the
+                    // last
+                    // fetch.
+                    val ids = loaded.value.map { it.id }.toSet()
+                    expandedSessions = expandedSessions intersect ids
+                    subagentLoads = subagentLoads.filterKeys { it in ids }
+                    expandedSessions.forEach(::loadSubagents)
                    loaded
                } catch (e: ApiException) {
                    LoadState.failed(e)
@@ -127,6 +169,17 @@ fun SessionListScreen(
                                deleting = session.id in deleting,
                                onOpen = { onOpen(session) },
                                onLongPress = { confirmingDelete = session },
+                                expanded = session.id in expandedSessions,
+                                subagents = subagentLoads[session.id],
+                                onToggleSubagents = {
+                                    if (session.id in expandedSessions) {
+                                        expandedSessions = expandedSessions - session.id
+                                    } else {
+                                        expandedSessions = expandedSessions + session.id
+                                        loadSubagents(session.id)
+                                    }
+                                },
+                                onOpenSubagent = { subagent -> onOpenSubagent(session, subagent) },
                            )
                            Spacer(Modifier.height(12.dp))
                        }
@@ -225,7 +278,7 @@ fun SessionListScreen(
                                    deleteSession(settings, session.id, alsoDeleteForeign)
                                    // After it succeeded, not before: a refused delete leaves the
                                    // session exactly as it was, and its transcript with it.
-                                    transcriptCache.session(session.id).purge()
+                                    transcriptCache.session(TranscriptAddress(session.id)).purge()
                                }
                                // Only this row, and only what changed. Refetching the list instead
                                // put every other session back through loading and handed the
@@ -276,6 +329,12 @@ private fun SessionCard(
    deleting: Boolean,
    onOpen: () -> Unit,
    onLongPress: () -> Unit,
+    /** Whether the expander below is open. Collapsed by default; see [SessionListScreen]. */
+    expanded: Boolean,
+    /** What the expander's own fetch answered, or null before it has been asked. */
+    subagents: LoadState<List<SubagentSummary>>?,
+    onToggleSubagents: () -> Unit,
+    onOpenSubagent: (SubagentSummary) -> Unit,
 ) {
    BusyItem(label = if (deleting) "deleting" else null) {
        Card(
@@ -332,11 +391,105 @@ private fun SessionCard(
                        color = MaterialTheme.colorScheme.error,
                    )
                }
+                // Nothing at all for a card with no subagents: a disabled expander here would be
+                // noise on every ordinary session's card. Its own row at the bottom rather than
+                // beside the title or the machine line, so opening it never displaces text that was
+                // already on screen -- see UI_RULES on a control not displacing the text beside it.
+                if (session.subagents > 0) {
+                    Spacer(Modifier.height(8.dp))
+                    // The platform's minimum touch height, not the chevron's own ten or so dp:
+                    // at the chevron's height a tap meant for it landed on the first subcard
+                    // beneath and opened a subagent instead.
+                    Row(
+                        horizontalArrangement = Arrangement.Center,
+                        verticalAlignment = Alignment.CenterVertically,
+                        modifier =
+                            Modifier.fillMaxWidth()
+                                .heightIn(min = 48.dp)
+                                .clickable(enabled = !deleting, onClick = onToggleSubagents)
+                                .semantics {
+                                    contentDescription =
+                                        if (expanded) "Collapse subagents" else "Expand subagents"
+                                },
+                    ) {
+                        Chevron(if (expanded) Pointing.Up else Pointing.Down)
+                    }
+                    if (expanded) {
+                        Spacer(Modifier.height(4.dp))
+                        Column(verticalArrangement = Arrangement.spacedBy(8.dp)) {
+                            when (subagents) {
+                                null,
+                                is LoadState.Loading ->
+                                    CircularProgressIndicator(
+                                        modifier = Modifier.width(20.dp).height(20.dp),
+                                        strokeWidth = 2.dp,
+                                    )
+                                is LoadState.Error ->
+                                    // Said here rather than left silent: a fetch that failed and an
+                                    // expander that simply found nothing must not look the same --
+                                    // see UI_RULES on designing the unknown state first.
+                                    Text(
+                                        subagents.message,
+                                        style = MaterialTheme.typography.bodySmall,
+                                        color = MaterialTheme.colorScheme.error,
+                                    )
+                                is LoadState.Loaded ->
+                                    subagents.value.forEach { subagent ->
+                                        SubagentCard(
+                                            subagent,
+                                            onClick = { onOpenSubagent(subagent) },
+                                        )
+                                    }
+                            }
+                        }
+                    }
+                }
            }
        }
    }
 }

+/**
+ * One subagent, indented inside its session's card -- the way dev-updater draws a project's
+ * components (`ComponentCard`, `UpdaterScreen.kt`): an outlined card, not the session card's own
+ * filled one, so the nesting reads as one step rather than as another session.
+ */
+@Composable
+private fun SubagentCard(subagent: SubagentSummary, onClick: () -> Unit) {
+    OutlinedCard(Modifier.fillMaxWidth().clickable(onClick = onClick)) {
+        Column(Modifier.padding(horizontal = 12.dp, vertical = 8.dp)) {
+            Text(subagent.title, style = MaterialTheme.typography.titleSmall)
+            Spacer(Modifier.height(2.dp))
+            Row(modifier = Modifier.fillMaxWidth()) {
+                Text(
+                    subagentStatusLabel(subagent.status),
+                    style = MaterialTheme.typography.bodySmall,
+                    color = MaterialTheme.colorScheme.onSurfaceVariant,
+                    modifier = Modifier.weight(1f),
+                )
+                Text(
+                    relativeTime(subagent.lastActivity),
+                    style = MaterialTheme.typography.bodySmall,
+                    color = MaterialTheme.colorScheme.onSurfaceVariant,
+                )
+            }
+        }
+    }
+}
+
+/**
+ * The subcard's word for a subagent's status -- see SUBAGENTS.md's "Wire shape". Its own function
+ * rather than a branch inside [StatusText], because a subagent's three states are not that
+ * composable's five: "exited" reads as "finished" here, since its process was always its parent's
+ * and never something of its own to have merely stopped.
+ */
+private fun subagentStatusLabel(status: String) =
+    when (status) {
+        "running" -> "running"
+        "exited" -> "finished"
+        else -> "unknown"
+    }
+
@Composable
 fun StatusText(status: String) {
    val (label, color) =
@@ -216,16 +216,33 @@ fun SessionScreen(
    share: ShareRequest? = null,
    /** Said once [share] has been attached here, so it is not attached again. */
    onShareTaken: () -> Unit = {},
+    /**
+     * Draws this screen read-only, on a subagent's own transcript instead of the session's.
+     *
+     * A subagent has no process and no controls of its own -- see SUBAGENTS.md's "Phone" -- so
+     * every gate below keyed on this switches off the composer, the files button, the settings cog,
+     * the usage bar and notifications, while everything that draws a transcript (paging, cache,
+     * selection, images, the status row, stream reconnects) is reused unchanged, pointed at
+     * [address] instead of the session's own.
+     */
+    subagent: SubagentSummary? = null,
 ) {
    DebugStats.count("session screen recomposed")
+    val isSubagent = subagent != null
+    val address = TranscriptAddress(summary.id, subagent?.id)
    val scope = rememberCoroutineScope()
    val topEdgeHeld = remember { TopEdgeHold() }
    var items by remember { mutableStateOf(listOf<TranscriptItem>()) }
-    var status by remember { mutableStateOf(summary.status) }
+    var status by remember { mutableStateOf(subagent?.status ?: summary.status) }
    // Seeded from the row this screen was opened from, so a conversation already under way says how
    // much it is holding before any turn happens here. Null is "nobody has measured it", which is a
    // different answer from an empty context and is drawn differently.
-    var contextTokens by remember(summary.id) { mutableStateOf(summary.contextTokens) }
+    //
+    // A subagent has no context measurement of its own, so it always starts unmeasured rather than
+    // borrowing the parent session's figure -- see UI_RULES on not showing an inferred value as one
+    // that was measured.
+    var contextTokens by
+        remember(address) { mutableStateOf(if (isSubagent) null else summary.contextTokens) }
    // When the current compaction started. The moment comes off the `compacting` status event
    // itself -- the server timestamps every transcript line -- rather than off this device noticing
    // one, which is what makes it survive leaving the session and reopening it.
@@ -241,7 +258,13 @@ fun SessionScreen(
    val context = LocalContext.current
    // Seeded from what was left in the box last time and written back on every keystroke, so
    // leaving the screen does not throw away a half-typed message. See `Drafts.kt`.
-    var input by remember(summary.id) { mutableStateOf(atEnd(loadDraft(context, summary.id))) }
+    //
+    // A subagent has no box to type into, so it never touches a draft at all -- not this session's,
+    // which is what reading one keyed only by `summary.id` would do here.
+    var input by
+        remember(summary.id) {
+            mutableStateOf(if (isSubagent) atEnd("") else atEnd(loadDraft(context, summary.id)))
+        }
    // A model the reader has chosen and not yet confirmed. See [ModelSwitchWarning]: switching
    // makes the session re-read the whole conversation.
    var pendingModel by remember { mutableStateOf<String?>(null) }
@@ -294,27 +317,26 @@ fun SessionScreen(
    // Reload throws away what it was reading from.
    val cache = remember(settings) { TranscriptCache(cacheRoot(context, settings)) }
    val source =
-        remember(summary.id, epoch) {
-            TranscriptSource(settings, summary.id, cache.session(summary.id))
-        }
+        remember(address, epoch) { TranscriptSource(settings, address, cache.session(address)) }
    // Whether the cached tail has been shown to still be the server's own line. Nothing is resumed
    // from a cached cursor until it has, and a probe that could not be made leaves this false for
    // the stream loop to try again.
-    var probePassed by remember(summary.id, epoch) { mutableStateOf(false) }
+    var probePassed by remember(address, epoch) { mutableStateOf(false) }
    // Whether the opening effect is still settling that question. It draws the cached rows and
    // lifts [ready] before the answer arrives, which is the point of the cache -- so the stream
    // below waits for this rather than for `ready`, or it asks the same question twice.
-    var probing by remember(summary.id, epoch) { mutableStateOf(true) }
+    var probing by remember(address, epoch) { mutableStateOf(true) }
    // The oldest sequence number loaded, and whether there is more behind it. Paging backwards is
    // what keeps opening a long session cheap.
    var oldestSeq by remember { mutableLongStateOf(0L) }
-    // Where this session was last being read, from this device's own store. Read once, because the
-    // answer stops being interesting the moment the list is on screen.
-    val savedAnchor = remember(summary.id, epoch) { loadScrollAnchor(context, summary.id) }
+    // Where this transcript was last being read, from this device's own store, keyed by the address
+    // rather than the session id so a subagent's saved position cannot collide with its session's.
+    // Read once, because the answer stops being interesting the moment the list is on screen.
+    val savedAnchor = remember(address, epoch) { loadScrollAnchor(context, address.cachePath) }
    // Whether the saved position is still being put back. Nothing is drawn while it is: opening at
    // the newest end and then travelling to the anchor is exactly the journey a reader must never
    // see.
-    var restoring by remember(summary.id, epoch) { mutableStateOf(savedAnchor != null) }
+    var restoring by remember(address, epoch) { mutableStateOf(savedAnchor != null) }
    // Messages the server has taken and the session has not read yet, by the id that will resolve
    // them. From the event stream rather than from what this screen sent, so they survive leaving
    // the session -- and a message sent from another device is drawn waiting on this one too.
@@ -327,11 +349,11 @@ fun SessionScreen(
    var loadingHistory by remember { mutableStateOf(false) }
    var ready by remember { mutableStateOf(false) }
    // Replies parsed ahead of the rows that draw them; see [ParsedReplies].
-    val replies = remember(summary.id) { ParsedReplies() }
-    // Keyed like everything else describing one session's transcript. `rememberLazyListState` saves
-    // through `rememberSaveable`, and this screen restores by its own anchor instead -- two
-    // restores would fight over the first frame.
-    val listState = remember(summary.id) { LazyListState() }
+    val replies = remember(address) { ParsedReplies() }
+    // Keyed like everything else describing one transcript. `rememberLazyListState` saves through
+    // `rememberSaveable`, and this screen restores by its own anchor instead -- two restores would
+    // fight over the first frame.
+    val listState = remember(address) { LazyListState() }
    // Whether the newest message is on screen right now. The list is reversed, so the newest end is
    // the scrolling start: nothing behind you is exactly being at the bottom. Asked of the scroll
    // state rather than of item indices, because a zero-height first item makes an index ambiguous.
@@ -637,7 +659,7 @@ fun SessionScreen(
    // ended and carries live events only. The window comes from this phone's own copy when there is
    // one, and then costs a single request to check that the server's transcript is still the one
    // it came from. See TRANSCRIPT_CACHE.md.
-    LaunchedEffect(summary.id, epoch) {
+    LaunchedEffect(address, epoch) {
        /**
         * One opening window onto the screen, whichever side it came from.
         *
@@ -667,11 +689,16 @@ fun SessionScreen(
                // A replay is as old as the last visit; the row this screen was opened from was
                // fetched moments ago. So the transcript comes from the cache and everything that
                // is not the transcript comes from the summary -- otherwise a session that finished
-                // an hour ago opens saying "working" until the stream connects.
-                status = summary.status
-                model = summary.model
-                permissionMode = summary.permissionMode ?: "auto"
-                if (summary.status != "compacting") compactingSince = null
+                // an hour ago opens saying "working" until the stream connects. A subagent's status
+                // comes from its own summary, never the parent session's: they are two different
+                // things running or not, and the parent's model and permission mode do not apply to
+                // it at all.
+                status = subagent?.status ?: summary.status
+                if (!isSubagent) {
+                    model = summary.model
+                    permissionMode = summary.permissionMode ?: "auto"
+                }
+                if (status != "compacting") compactingSince = null
                // Nothing to put back, so these rows are the screen and the probe can return under
                // them. A restore still has history to fetch and is gated below.
                if (savedAnchor == null) ready = true
@@ -798,7 +825,7 @@ fun SessionScreen(
    // at the top on their return. Switching apps is a choice somebody made, not a fault to report.
    // Stopping the stream deliberately makes the drop a close rather than an error, and resuming
    // reconnects from the same cursor.
-    LaunchedEffect(summary.id, ready, epoch, lifecycleOwner) {
+    LaunchedEffect(address, ready, epoch, lifecycleOwner) {
        if (!ready) return@LaunchedEffect
        // The opening effect draws cached rows and lifts `ready` *before* it has checked that the
        // cursor under them is still the server's, so `ready` is no longer the whole gate. Without
@@ -868,17 +895,22 @@ fun SessionScreen(
    // The screen going away entirely, which the lifecycle scope above does not cover: a composable
    // can leave the composition while the activity stays started. Keyed on the epoch as well, so
    // Reload's replacement source is the one a later disposal closes.
-    DisposableEffect(summary.id, epoch) { onDispose { source.close() } }
+    DisposableEffect(address, epoch) { onDispose { source.close() } }

    // Nothing gets announced about the session somebody is reading; see NotificationService.
    // RESUMED rather than STARTED because "looking at it" means the foreground.
-    LaunchedEffect(summary.id, lifecycleOwner) {
-        lifecycleOwner.repeatOnLifecycle(Lifecycle.State.RESUMED) {
-            NotificationService.showing(context, summary.id)
-            try {
-                awaitCancellation()
-            } finally {
-                NotificationService.stoppedShowing(summary.id)
+    //
+    // Not for a subagent: it has no notifications of its own, and it is not the session this would
+    // otherwise mark as being read.
+    if (!isSubagent) {
+        LaunchedEffect(summary.id, lifecycleOwner) {
+            lifecycleOwner.repeatOnLifecycle(Lifecycle.State.RESUMED) {
+                NotificationService.showing(context, summary.id)
+                try {
+                    awaitCancellation()
+                } finally {
+                    NotificationService.stoppedShowing(summary.id)
+                }
            }
        }
    }
@@ -924,7 +956,7 @@ fun SessionScreen(
                val (index, offset, awayFromNewest) = settled
                saveScrollAnchor(
                    context,
-                    summary.id,
+                    address.cachePath,
                    // Nothing to restore at the newest end, which is where a session with no anchor
                    // opens anyway. One *before* the index, because item zero is the "below" slot.
                    if (!awayFromNewest) null
@@ -947,7 +979,7 @@ fun SessionScreen(
    //
    // There is no correction beside this one. Following the newest message is not an effect: the
    // list is reversed, so an arriving message extends the end the viewport is pinned to.
-    val unitSizes = remember(summary.id) { HashMap<Any, Int>() }
+    val unitSizes = remember(address) { HashMap<Any, Int>() }
    LaunchedEffect(listState, moreHistory) {
        snapshotFlow { listState.layoutInfo }
            .collect { info ->
@@ -983,21 +1015,25 @@ fun SessionScreen(
            }
    }

-    LaunchedEffect(summary.setupName, summary.provider) {
-        offeredModels =
-            try {
-                withContext(Dispatchers.IO) {
-                    fetchSetups(settings)
-                        .firstOrNull { it.name == summary.setupName }
-                        ?.providers
-                        ?.firstOrNull { it.name == summary.provider }
-                        ?.models
-                        .orEmpty()
+    // Only for the model picker, which a subagent does not have.
+    if (!isSubagent) {
+        LaunchedEffect(summary.setupName, summary.provider) {
+            offeredModels =
+                try {
+                    withContext(Dispatchers.IO) {
+                        fetchSetups(settings)
+                            .firstOrNull { it.name == summary.setupName }
+                            ?.providers
+                            ?.firstOrNull { it.name == summary.provider }
+                            ?.models
+                            .orEmpty()
+                    }
+                } catch (_: Exception) {
+                    // Not worth reporting: the picker simply has nothing to offer, which is
+                    // visible.
+                    emptyList()
                }
-            } catch (_: Exception) {
-                // Not worth reporting: the picker simply has nothing to offer, which is visible.
-                emptyList()
-            }
+        }
    }

    /**
@@ -1148,8 +1184,9 @@ fun SessionScreen(
    }

    // One poll for the machines' limits, read by everything on this screen that reports them.
-    val usageFeed = rememberUsageFeed(settings)
-    val usage = usageFeed.forSetup(summary.setup)
+    // Nothing meters a subagent -- it has no account of its own -- so it never starts this poll.
+    val usageFeed = if (isSubagent) null else rememberUsageFeed(settings)
+    val usage = usageFeed?.forSession(summary) ?: SessionUsage.NotMetered
    RecordFrames()
    var usageOpen by remember { mutableStateOf(false) }
    var settingsOpen by remember { mutableStateOf(false) }
@@ -1191,7 +1228,12 @@ fun SessionScreen(
    // the bench scripts keep working when this moves again. They pressed it at a hand-measured
    // coordinate until 2026-09-03, and anything that moved the header made that tap land on
    // whatever now sat there -- reporting a number that was never measured.
-    val copyRenderReport = {
+    // Shared by the ordinary "Copy" button and (bench build only) "Run benchmark": what differs
+    // between them is only whether there is a [extra] section, built by BenchRun.run beforehand --
+    // everything about assembling, copying and logging the report is exactly the same act either
+    // way, and a second copy of it beside `onRunBenchmark` below would be the two silently
+    // disagreeing about what "the report" contains the first time either one changed.
+    fun buildAndCopyReport(extra: List<String> = emptyList()) {
        val report =
            debugReport(
                device =
@@ -1213,6 +1255,7 @@ fun SessionScreen(
                accounting =
                    FrameStats.drawPhase().let { (nanos, count) -> drawAccounting(nanos, count) },
                crash = lastCrash(context),
+                extra = extra,
            )
        context.copyToClipboard("ai-app render report", report)
        // Also to the log, so a session driving the app over adb can read the same report the
@@ -1226,6 +1269,20 @@ fun SessionScreen(
        DebugStats.reset()
        Toast.makeText(context, "Copied render report", Toast.LENGTH_SHORT).show()
    }
+    val copyRenderReport = { buildAndCopyReport() }
+    // Bench build only: P0's scripted scroll-and-stream benchmark (BenchRun.kt), against the
+    // fixture session opened below instead of a real server. Null everywhere else -- see
+    // [SessionSettingsDialog]'s onRunBenchmark.
+    val runBenchmark: (() -> Unit)? =
+        if (BuildConfig.FIXTURE_MODE) {
+            {
+                settingsOpen = false
+                scope.launch {
+                    val extra = BenchRun.run(context, scope, listState)
+                    buildAndCopyReport(extra)
+                }
+            }
+        } else null
    Box(Modifier.fillMaxSize()) {
        Column(Modifier.fillMaxSize()) {
            Row(
@@ -1236,20 +1293,35 @@ fun SessionScreen(
                // A ring's worth, which is what the arrow already keeps on its other three sides.
                Spacer(Modifier.width(GLYPH_BUTTON_MARGIN))
                Column(Modifier.weight(1f)) {
-                    Text(title, style = MaterialTheme.typography.titleMedium)
-                    // Machine first, then what runs on it -- the same order and the same wording
-                    // everywhere this pair appears, so it reads as one fact rather than two
-                    // sentences with different grammar.
-                    //
-                    // No model. The picker in the footer already shows what this session is set to,
-                    // and showing it twice means two things to keep in step -- they disagreed for a
-                    // moment on every model change.
-                    Text(
-                        "${summary.setupName} · ${summary.provider}",
-                        style = MaterialTheme.typography.bodySmall,
-                        color = MaterialTheme.colorScheme.onSurfaceVariant,
-                    )
+                    // A subagent's own title, with the session's beneath it in a smaller style --
+                    // the header says whose conversation this is as well as what it is. Otherwise
+                    // just the session's title, as before.
+                    if (subagent != null) {
+                        Text(subagent.title, style = MaterialTheme.typography.titleMedium)
+                        Text(
+                            title,
+                            style = MaterialTheme.typography.bodySmall,
+                            color = MaterialTheme.colorScheme.onSurfaceVariant,
+                        )
+                    } else {
+                        Text(title, style = MaterialTheme.typography.titleMedium)
+                        // Machine first, then what runs on it -- the same order and the same
+                        // wording everywhere this pair appears, so it reads as one fact rather than
+                        // two sentences with different grammar.
+                        //
+                        // No model. The picker in the footer already shows what this session is set
+                        // to, and showing it twice means two things to keep in step -- they
+                        // disagreed for a moment on every model change.
+                        Text(
+                            "${summary.setupName} · ${summary.provider}",
+                            style = MaterialTheme.typography.bodySmall,
+                            color = MaterialTheme.colorScheme.onSurfaceVariant,
+                        )
+                    }
                }
+                // None of this is a subagent's: it has no files of its own to browse, no settings,
+                // and nothing meters it -- see SUBAGENTS.md's "Phone".
+                //
                // Beside the provider it reports on, which is the line directly to its left. Its
                // real home is this provider's settings, which do not exist yet. A session on a
                // provider with no such service gets an honest "unavailable" rather than a hidden
@@ -1264,42 +1336,47 @@ fun SessionScreen(
                // Usage, files, settings -- widest scope first, narrowing to the right, so the cog
                // stays at the end where every other screen keeps it. Asked for in this order by
                // Iris on 2026-09-03.
-                Row {
-                    GlyphButton(
-                        USAGE_GLYPH,
-                        "Usage",
-                        { usageOpen = true },
-                        colour = usageGlyphColour(usage),
-                    )
-                    // The machine's files, which is where the answer to "what did it actually
-                    // change" is. It opens *over* this screen rather than replacing it.
-                    GlyphButton(
-                        FOLDER_GLYPH,
-                        "Files",
-                        onClick = {
-                            onFiles(
-                                FilesTarget(
-                                    setup = summary.setup,
-                                    setupName = summary.setupName,
-                                    // Where this session works, and the machine's own home when it
-                                    // was never given a directory -- resolved there rather than
-                                    // guessed at here, since this app does not know that home.
-                                    start = summary.cwd?.takeIf { it.isNotBlank() } ?: "~",
+                if (!isSubagent) {
+                    Row {
+                        GlyphButton(
+                            USAGE_GLYPH,
+                            "Usage",
+                            { usageOpen = true },
+                            colour = usageGlyphColour(usage),
+                        )
+                        // The machine's files, which is where the answer to "what did it actually
+                        // change" is. It opens *over* this screen rather than replacing it.
+                        GlyphButton(
+                            FOLDER_GLYPH,
+                            "Files",
+                            onClick = {
+                                onFiles(
+                                    FilesTarget(
+                                        setup = summary.setup,
+                                        setupName = summary.setupName,
+                                        // Where this session works, and the machine's own home when
+                                        // it was never given a directory -- resolved there rather
+                                        // than guessed at here, since this app does not know that
+                                        // home.
+                                        start = summary.cwd?.takeIf { it.isNotBlank() } ?: "~",
+                                    )
                                )
-                            )
-                        },
-                    )
-                    // What it opens is about this session, so it sits at the end of the session's
-                    // own row. A cog and not a word because there will be more, and a bar of words
-                    // has nowhere to put it.
-                    GlyphButton(SETTINGS_GLYPH, "Session settings", { settingsOpen = true })
+                            },
+                        )
+                        // What it opens is about this session, so it sits at the end of the
+                        // session's own row. A cog and not a word because there will be more, and a
+                        // bar of words has nowhere to put it.
+                        GlyphButton(SETTINGS_GLYPH, "Session settings", { settingsOpen = true })
+                    }
                }
            }

            // Under the header, above everything the session itself says: it is a fact about the
            // machine rather than a turn in the conversation, and it is the number that decides
-            // whether to keep going.
-            SessionUsageBar(usage)
+            // whether to keep going. Nothing meters a subagent.
+            if (!isSubagent) {
+                SessionUsageBar(usage)
+            }

            (streamError ?: actionError)?.let { message ->
                Text(
@@ -1545,6 +1622,7 @@ fun SessionScreen(
                                                is TranscriptItem.ClearedNote -> ClearedRow()
                                                is TranscriptItem.CompactedNote ->
                                                    CompactedRow(item)
+                                                is TranscriptItem.LimitNote -> LimitRow(item)
                                                // Never reached: a peer message is flattened into
                                                // its own units. Here because a `when` over the
                                                // item kinds has to stay exhaustive.
@@ -1643,185 +1721,208 @@ fun SessionScreen(
                )
            }

+            // Kept for a subagent -- see SUBAGENTS.md's "Phone" -- with the wording that turns
+            // "exited" into "finished" for one, since it has no process to leave running or stop.
            SessionStatusRow(
                status = status,
                compactingFor = compactingFor,
                contextTokens = contextTokens,
+                subagent = isSubagent,
            )

-            // Between the transcript and the box: above what is being typed, so the list does not
-            // cover the thing the command is about, and below everything that explains it.
-            CommandSuggestions(
-                // Nothing to suggest about a suggestion that was just taken. `/compact` is a whole
-                // command *and* a prefix of itself, so picking it left the list standing there with
-                // the one row already chosen. Held by what was picked rather than by a flag, so
-                // typing anything else brings the list back without a second thing to reset.
-                commands = if (input.text == picked) emptyList() else suggestedCommands(input.text),
-                onPick = { command ->
-                    // At the end of what was inserted, which is where the reader carries on typing:
-                    // a command with an argument is put in the box half-written, and a cursor left
-                    // at the front makes the next keystroke the first character of "/rename".
-                    input = atEnd(command.typed())
-                    picked = command.typed()
-                },
-            )
-
-            // Always enabled -- a send while the session is running becomes a steering message
-            // injected at the next tool boundary, which is the point of the whole app.
-            //
-            // The field gets a row of its own, above the buttons: sharing one put the full width
-            // behind three controls, so the thing being typed into was the narrowest on the row.
-            Column(Modifier.fillMaxWidth().padding(8.dp)) {
-                // Directly above the box they will be sent from, so what is attached is visible
-                // rather than counted: the "+2" on the button below said how many and never which.
-                PendingAttachments(
-                    settings = settings,
-                    sessionId = summary.id,
-                    refs = pendingAttachments,
-                    onRemove = { pendingAttachments = pendingAttachments - it },
-                )
-                OutlinedTextField(
-                    value = input,
-                    onValueChange = {
-                        input = it
-                        saveDraft(context, summary.id, it.text)
+            // Everything from here down is the composer: a subagent cannot be messaged, so none of
+            // it applies -- see SUBAGENTS.md's "Phone".
+            if (!isSubagent) {
+                // Between the transcript and the box: above what is being typed, so the list does
+                // not cover the thing the command is about, and below everything that explains it.
+                CommandSuggestions(
+                    // Nothing to suggest about a suggestion that was just taken. `/compact` is a
+                    // whole command *and* a prefix of itself, so picking it left the list standing
+                    // there with the one row already chosen. Held by what was picked rather than by
+                    // a flag, so typing anything else brings the list back without a second thing
+                    // to
+                    // reset.
+                    commands =
+                        if (input.text == picked) emptyList() else suggestedCommands(input.text),
+                    onPick = { command ->
+                        // At the end of what was inserted, which is where the reader carries on
+                        // typing: a command with an argument is put in the box half-written, and a
+                        // cursor left at the front makes the next keystroke the first character of
+                        // "/rename".
+                        input = atEnd(command.typed())
+                        picked = command.typed()
                    },
-                    modifier = Modifier.fillMaxWidth(),
-                    // No longer "(+image)": the images are on screen above this, and a placeholder
-                    // saying so said it in words beside the thing itself.
-                    placeholder = { Text("Message") },
-                    maxLines = 4,
                )
-                Row(
-                    verticalAlignment = Alignment.CenterVertically,
-                    modifier = Modifier.fillMaxWidth(),
-                ) {
-                    // Photo or file, asked here rather than by two buttons: the row is full, and
-                    // attaching is one action whichever picker answers it.
-                    var attaching by remember { mutableStateOf(false) }
-                    Box {
-                        // Just "+". The count it used to carry was standing in for showing them.
-                        BubbleButton(onClick = { attaching = true }) { Text("+") }
-                        DropdownMenu(
-                            expanded = attaching,
-                            onDismissRequest = { attaching = false },
-                            // See PickerButton: without this the menu opens a status bar's height
-                            // away from the button in an edge-to-edge activity.
-                            properties = PopupProperties(clippingEnabled = false),
-                            shape = BubbleMenuShape,
-                        ) {
-                            DropdownMenuItem(
-                                text = { Text("Photo") },
-                                onClick = {
-                                    attaching = false
-                                    pickImage.launch(
-                                        PickVisualMediaRequest(
-                                            ActivityResultContracts.PickVisualMedia.ImageOnly
-                                        )
-                                    )
-                                },
-                            )
-                            DropdownMenuItem(
-                                text = { Text("File") },
-                                onClick = {
-                                    attaching = false
-                                    pickFile.launch(arrayOf("*/*"))
-                                },
-                            )
-                        }
-                    }
-                    // The settings share what is left after the actions have taken what they need.
-                    // A Row hands out intrinsic widths in order and clips whatever runs past the
-                    // edge, so with these laid out first the arrival of Stop pushed Send off the
-                    // screen entirely -- the app's central control, gone at the moment it is most
-                    // in use.
+
+                // Always enabled -- a send while the session is running becomes a steering message
+                // injected at the next tool boundary, which is the point of the whole app.
+                //
+                // The field gets a row of its own, above the buttons: sharing one put the full
+                // width
+                // behind three controls, so the thing being typed into was the narrowest on the
+                // row.
+                Column(Modifier.fillMaxWidth().padding(8.dp)) {
+                    // Directly above the box they will be sent from, so what is attached is visible
+                    // rather than counted: the "+2" on the button below said how many and never
+                    // which.
+                    PendingAttachments(
+                        settings = settings,
+                        sessionId = summary.id,
+                        refs = pendingAttachments,
+                        onRemove = { pendingAttachments = pendingAttachments - it },
+                    )
+                    OutlinedTextField(
+                        value = input,
+                        onValueChange = {
+                            input = it
+                            saveDraft(context, summary.id, it.text)
+                        },
+                        modifier = Modifier.fillMaxWidth(),
+                        // No longer "(+image)": the images are on screen above this, and a
+                        // placeholder saying so said it in words beside the thing itself.
+                        placeholder = { Text("Message") },
+                        maxLines = 4,
+                    )
                    Row(
                        verticalAlignment = Alignment.CenterVertically,
-                        modifier = Modifier.weight(1f),
+                        modifier = Modifier.fillMaxWidth(),
                    ) {
-                        if (offeredModels.isNotEmpty()) {
+                        // Photo or file, asked here rather than by two buttons: the row is full,
+                        // and
+                        // attaching is one action whichever picker answers it.
+                        var attaching by remember { mutableStateOf(false) }
+                        Box {
+                            // Just "+". The count it used to carry was standing in for showing
+                            // them.
+                            BubbleButton(onClick = { attaching = true }) { Text("+") }
+                            DropdownMenu(
+                                expanded = attaching,
+                                onDismissRequest = { attaching = false },
+                                // See PickerButton: without this the menu opens a status bar's
+                                // height away from the button in an edge-to-edge activity.
+                                properties = PopupProperties(clippingEnabled = false),
+                                shape = BubbleMenuShape,
+                            ) {
+                                DropdownMenuItem(
+                                    text = { Text("Photo") },
+                                    onClick = {
+                                        attaching = false
+                                        pickImage.launch(
+                                            PickVisualMediaRequest(
+                                                ActivityResultContracts.PickVisualMedia.ImageOnly
+                                            )
+                                        )
+                                    },
+                                )
+                                DropdownMenuItem(
+                                    text = { Text("File") },
+                                    onClick = {
+                                        attaching = false
+                                        pickFile.launch(arrayOf("*/*"))
+                                    },
+                                )
+                            }
+                        }
+                        // The settings share what is left after the actions have taken what they
+                        // need. A Row hands out intrinsic widths in order and clips whatever runs
+                        // past the edge, so with these laid out first the arrival of Stop pushed
+                        // Send off the screen entirely -- the app's central control, gone at the
+                        // moment it is most in use.
+                        Row(
+                            verticalAlignment = Alignment.CenterVertically,
+                            modifier = Modifier.weight(1f),
+                        ) {
+                            if (offeredModels.isNotEmpty()) {
+                                PickerButton(
+                                    current = modelLabel(model),
+                                    // What the machine offers, plus the state a session is in when
+                                    // it has chosen none of them. The button has always been able
+                                    // to
+                                    // say "default"; until this the list could not, so leaving it
+                                    // was a one-way trip.
+                                    options = listOf(DEFAULT_MODEL) + offeredModels,
+                                    // Not set here. The button follows what the session reports it
+                                    // is set to, which arrives a moment later and is sometimes a
+                                    // different answer -- a name the CLI resolved, or no change at
+                                    // all on a provider whose model is fixed. Asked about first,
+                                    // unless there is nothing to lose by it -- see
+                                    // [ModelSwitchWarning].
+                                    onPick = { chosen ->
+                                        if (
+                                            modelLabel(chosen) == modelLabel(model) ||
+                                                !worthWarningAbout(status, contextTokens, items)
+                                        ) {
+                                            act { setSessionModel(settings, summary.id, chosen) }
+                                        } else {
+                                            pendingModel = chosen
+                                        }
+                                    },
+                                )
+                            }
                            PickerButton(
-                                current = modelLabel(model),
-                                // What the machine offers, plus the state a session is in when it
-                                // has chosen none of them. The button has always been able to say
-                                // "default"; until this the list could not, so leaving it was a
-                                // one-way trip.
-                                options = listOf(DEFAULT_MODEL) + offeredModels,
-                                // Not set here. The button follows what the session reports it is
-                                // set to, which arrives a moment later and is sometimes a different
-                                // answer -- a name the CLI resolved, or no change at all on a
-                                // provider whose model is fixed. Asked about first, unless there is
-                                // nothing to lose by it -- see [ModelSwitchWarning].
+                                current = permissionMode,
+                                options = PERMISSION_MODES,
                                onPick = { chosen ->
-                                    if (
-                                        modelLabel(chosen) == modelLabel(model) ||
-                                            !worthWarningAbout(status, contextTokens, items)
-                                    ) {
-                                        act { setSessionModel(settings, summary.id, chosen) }
-                                    } else {
-                                        pendingModel = chosen
-                                    }
+                                    act { setSessionPermissionMode(settings, summary.id, chosen) }
                                },
                            )
                        }
-                        PickerButton(
-                            current = permissionMode,
-                            options = PERMISSION_MODES,
-                            onPick = { chosen ->
-                                act { setSessionPermissionMode(settings, summary.id, chosen) }
-                            },
-                        )
-                    }
-                    // The same filled shape as the button beside it, not an outlined one: these are
-                    // two things you can do about the session, and weighting one as secondary said
-                    // they were a primary action and its qualifier. What separates them is the
-                    // colour and the mark, which is what they mean.
-                    //
-                    // Always here, rather than arriving with the turn as it used to. A control that
-                    // comes and goes makes its own presence the signal, and a button always in the
-                    // same place also cannot push Send off the end of the row by turning up.
-                    val process =
-                        when {
-                            running -> ProcessAction.Pause
-                            status == "exited" -> ProcessAction.Start
-                            else -> ProcessAction.Stop
-                        }
-                    Button(
-                        onClick = {
-                            processInFlight = true
-                            act(onDone = { processInFlight = false }) {
-                                process.perform(settings, summary.id)
+                        // The same filled shape as the button beside it, not an outlined one: these
+                        // are two things you can do about the session, and weighting one as
+                        // secondary said they were a primary action and its qualifier. What
+                        // separates them is the colour and the mark, which is what they mean.
+                        //
+                        // Always here, rather than arriving with the turn as it used to. A control
+                        // that comes and goes makes its own presence the signal, and a button
+                        // always
+                        // in the same place also cannot push Send off the end of the row by turning
+                        // up.
+                        val process =
+                            when {
+                                running -> ProcessAction.Pause
+                                status == "exited" -> ProcessAction.Start
+                                else -> ProcessAction.Stop
                            }
-                        },
-                        enabled = !processInFlight,
-                        colors = actionButtonColors(process.colour()),
-                    ) {
-                        Glyph(
-                            process.glyph,
-                            colour = LocalContentColor.current,
-                            modifier = Modifier.semantics { contentDescription = process.label },
-                        )
-                    }
-                    Spacer(Modifier.width(8.dp))
-                    // The paper plane, with a clock on it while a turn is in flight: sending then
-                    // queues the message for the next tool boundary rather than starting a turn of
-                    // its own, and the two have to be told apart at a glance. The label says the
-                    // same thing to a screen reader.
-                    //
-                    // Disabled while there is nothing to send, rather than pressable and silent:
-                    // `send` has always returned early on an empty composer, so the button promised
-                    // something it would not do. Disabled and not hidden, for the reason above.
-                    Button(
-                        onClick = { send() },
-                        enabled = input.text.isNotBlank() || pendingAttachments.isNotEmpty(),
-                        colors = actionButtonColors(if (running) queueColor else sendColor),
-                    ) {
-                        Glyph(
-                            if (running) QUEUE_GLYPH else SEND_GLYPH,
-                            colour = LocalContentColor.current,
-                            modifier =
-                                Modifier.semantics { contentDescription = sendLabel(running) },
-                        )
+                        Button(
+                            onClick = {
+                                processInFlight = true
+                                act(onDone = { processInFlight = false }) {
+                                    process.perform(settings, summary.id)
+                                }
+                            },
+                            enabled = !processInFlight,
+                            colors = actionButtonColors(process.colour()),
+                        ) {
+                            Glyph(
+                                process.glyph,
+                                colour = LocalContentColor.current,
+                                modifier =
+                                    Modifier.semantics { contentDescription = process.label },
+                            )
+                        }
+                        Spacer(Modifier.width(8.dp))
+                        // The paper plane, with a clock on it while a turn is in flight: sending
+                        // then queues the message for the next tool boundary rather than starting a
+                        // turn of its own, and the two have to be told apart at a glance. The label
+                        // says the same thing to a screen reader.
+                        //
+                        // Disabled while there is nothing to send, rather than pressable and
+                        // silent:
+                        // `send` has always returned early on an empty composer, so the button
+                        // promised something it would not do. Disabled and not hidden, for the
+                        // reason above.
+                        Button(
+                            onClick = { send() },
+                            enabled = input.text.isNotBlank() || pendingAttachments.isNotEmpty(),
+                            colors = actionButtonColors(if (running) queueColor else sendColor),
+                        ) {
+                            Glyph(
+                                if (running) QUEUE_GLYPH else SEND_GLYPH,
+                                colour = LocalContentColor.current,
+                                modifier =
+                                    Modifier.semantics { contentDescription = sendLabel(running) },
+                            )
+                        }
                    }
                }
            }
@@ -1832,7 +1933,7 @@ fun SessionScreen(
    // is the screen's business rather than any row's. See [SessionImageViewer].
    fullImage?.let { ref -> SessionImageViewer(settings, summary.id, ref) { fullImage = null } }
    if (usageOpen) {
-        UsageDialog(feed = usageFeed, onDismiss = { usageOpen = false })
+        usageFeed?.let { UsageDialog(feed = it, onDismiss = { usageOpen = false }) }
    }
    if (settingsOpen) {
        // Measured when the dialog opens rather than kept up to date: what the reader is being told
@@ -1846,6 +1947,8 @@ fun SessionScreen(
            settings = settings,
            sessionId = summary.id,
            title = title,
+            effort = summary.effort.takeIf { summary.takesEffort },
+            takesEffort = summary.takesEffort,
            cachedBytes = cachedBytes,
            // The purge finishes before the epoch moves, because the relaunched opening effect
            // reads the same directory and would otherwise draw what is about to be deleted. The
@@ -1869,6 +1972,7 @@ fun SessionScreen(
            },
            onDismiss = { settingsOpen = false },
            onCopyRenderReport = copyRenderReport,
+            onRunBenchmark = runBenchmark,
        )
    }
 }
@@ -2122,6 +2226,13 @@ private fun SessionStatusRow(
    /** Context the session is holding, or null where nothing has measured it. */
    contextTokens: Long?,
    modifier: Modifier = Modifier,
+    /**
+     * Whether this row is for a subagent rather than a session, which changes only one word:
+     * "exited" reads as "finished" there too, the same as the subagent list's own card -- a
+     * subagent's process was always its parent's, so "exited" would read as a fault rather than the
+     * ordinary way one of these ends.
+     */
+    subagent: Boolean = false,
 ) {
    DebugStats.count("status row recomposed")
    Row(
@@ -2178,7 +2289,7 @@ private fun SessionStatusRow(
                Text(
                    when (status) {
                        "idle" -> "idle"
-                        "exited" -> "exited"
+                        "exited" -> if (subagent) "finished" else "exited"
                        "awaitingInput" -> "your turn"
                        "unknown" -> "can't tell"
                        else -> status
@@ -2249,7 +2360,7 @@ private const val ONE_TAP_MS = 250L
 * session is set to without spending a second line on saying it.
 */
@Composable
-private fun PickerButton(current: String, options: List<String>, onPick: (String) -> Unit) {
+fun PickerButton(current: String, options: List<String>, onPick: (String) -> Unit) {
    var open by remember { mutableStateOf(false) }
    // When an outside touch last closed the menu.
    //
@@ -6,8 +6,10 @@ import androidx.compose.foundation.layout.Spacer
 import androidx.compose.foundation.layout.fillMaxWidth
 import androidx.compose.foundation.layout.height
 import androidx.compose.foundation.layout.width
+import androidx.compose.foundation.rememberScrollState
 import androidx.compose.foundation.text.KeyboardActions
 import androidx.compose.foundation.text.KeyboardOptions
+import androidx.compose.foundation.verticalScroll
 import androidx.compose.material3.AlertDialog
 import androidx.compose.material3.CircularProgressIndicator
 import androidx.compose.material3.MaterialTheme
@@ -26,6 +28,10 @@ import androidx.compose.ui.Alignment
 import androidx.compose.ui.Modifier
 import androidx.compose.ui.text.input.ImeAction
 import androidx.compose.ui.unit.dp
+import java.time.Instant
+import java.time.ZoneId
+import java.time.format.DateTimeFormatter
+import java.time.format.FormatStyle
 import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.launch
 import kotlinx.coroutines.withContext
@@ -54,6 +60,16 @@ fun SessionSettingsDialog(
     */
    title: String,
    onRenamed: (String) -> Unit,
+    /**
+     * How hard the model thinks, as the session reports it, or null for the CLI's own default.
+     *
+     * Taken from the row this dialog was opened over rather than fetched, because unlike the
+     * notification switch there is nothing else that changes it: the level is this app's to set and
+     * the server does not resolve it into something else.
+     */
+    effort: String?,
+    /** Whether a level does anything here; the row is left out entirely where it does not. */
+    takesEffort: Boolean,
    /**
     * What this phone is holding of the conversation, or null while that is being measured -- see
     * the Reload row below, which is what would discard it.
@@ -66,9 +82,18 @@ fun SessionSettingsDialog(
     * measures is that screen's own state.
     */
    onCopyRenderReport: () -> Unit,
+    /**
+     * Runs P0's scripted scroll-and-stream benchmark and copies the extended report, or null on
+     * every build but `bench` -- see [BuildConfig.FIXTURE_MODE] and BenchRun.kt. Null rather than
+     * always-present-but-disabled: this has no meaning at all outside the bench build, and a
+     * control with nothing behind it on every other build is not a state worth drawing.
+     */
+    onRunBenchmark: (() -> Unit)? = null,
 ) {
    val scope = rememberCoroutineScope()
    var name by remember(sessionId) { mutableStateOf(title) }
+    var level by remember(sessionId) { mutableStateOf(effort) }
+    var effortError by remember { mutableStateOf<String?>(null) }
    var saving by remember { mutableStateOf(false) }
    var error by remember { mutableStateOf<String?>(null) }
    // Null until the server has been asked. The row this dialog was opened over is a snapshot of
@@ -77,6 +102,16 @@ fun SessionSettingsDialog(
    // and a spinner sits beside it, which is what not knowing looks like.
    var notify by remember(sessionId) { mutableStateOf<Boolean?>(null) }
    var notifyError by remember { mutableStateOf<String?>(null) }
+    // The same three-state shape the notification switch has, for the same reason: until the
+    // server has answered, the switch is disabled rather than showing a position nothing confirmed.
+    var autoResume by remember(sessionId) { mutableStateOf<Boolean?>(null) }
+    var resumeMessage by remember(sessionId) { mutableStateOf(DEFAULT_RESUME_MESSAGE) }
+    // When the server next intends to ask whether the limit has lifted, or null when nothing is
+    // waiting. Read once with everything else: it moves on the server's schedule, not this
+    // screen's, and a figure that redrew itself here would be this app re-measuring what it was
+    // told.
+    var resumeAt by remember(sessionId) { mutableStateOf<Double?>(null) }
+    var resumeError by remember { mutableStateOf<String?>(null) }
    // Where the session works. Null until the server has been asked, for the same reason the switch
    // above is. An empty answer is a session that was never given a directory, which is not the
    // same as one whose directory is unknown -- the field is only enabled once one of those is
@@ -90,6 +125,9 @@ fun SessionSettingsDialog(
        try {
            val fresh = withContext(Dispatchers.IO) { fetchSession(settings, sessionId) }
            notify = fresh.notify
+            autoResume = fresh.autoResume
+            resumeMessage = fresh.autoResumeMessage
+            resumeAt = fresh.resumeAt
            cwd = fresh.cwd.orEmpty()
            typedCwd = fresh.cwd.orEmpty()
        } catch (e: ApiException) {
@@ -97,6 +135,8 @@ fun SessionSettingsDialog(
            // instead of offering a position nothing confirmed.
            notifyError = e.message
            notify = null
+            resumeError = e.message
+            autoResume = null
        }
    }

@@ -125,6 +165,26 @@ fun SessionSettingsDialog(
        }
    }

+    /**
+     * Chooses a thinking level, which ends the process the old level was launched with.
+     *
+     * Put back if the request is refused, for the reason the notification switch below gives: a
+     * control that stays where it was put after a refusal is stating something untrue.
+     */
+    fun setEffort(chosen: String?) {
+        val was = level
+        level = chosen
+        effortError = null
+        scope.launch {
+            try {
+                withContext(Dispatchers.IO) { setSessionEffort(settings, sessionId, chosen) }
+            } catch (e: ApiException) {
+                level = was
+                effortError = e.message
+            }
+        }
+    }
+
    // Moved optimistically so the switch answers the finger that moved it, and put back if the
    // request is refused -- a switch that waits for a round trip reads as broken on a slow tunnel,
    // and one that stays moved after a refusal lies.
@@ -142,6 +202,39 @@ fun SessionSettingsDialog(
        }
    }

+    /**
+     * Turns auto-resume on or off, or changes what it would say.
+     *
+     * One request for both, because the server takes one: switching it on and typing the message
+     * are two halves of the same decision, and sending them separately would leave a moment where
+     * the session is armed with the old words.
+     *
+     * Put back if refused, like the notification switch. Turning it off also clears what was
+     * scheduled -- said here rather than only on the server, or the row would go on naming a time
+     * that no longer exists.
+     */
+    fun setAutoResume(on: Boolean, message: String) {
+        val wasOn = autoResume
+        val wasMessage = resumeMessage
+        val wasAt = resumeAt
+        autoResume = on
+        resumeMessage = message
+        if (!on) resumeAt = null
+        resumeError = null
+        scope.launch {
+            try {
+                withContext(Dispatchers.IO) {
+                    setSessionAutoResume(settings, sessionId, on, message)
+                }
+            } catch (e: ApiException) {
+                autoResume = wasOn
+                resumeMessage = wasMessage
+                resumeAt = wasAt
+                resumeError = e.message
+            }
+        }
+    }
+
    // Nothing to do when the name has not changed, so the button says so rather than sending a
    // request whose success would look exactly like the failure of having typed nothing.
    val changed = name.trim().isNotEmpty() && name.trim() != title
@@ -168,7 +261,10 @@ fun SessionSettingsDialog(
        onDismissRequest = onDismiss,
        title = { Text("Session settings") },
        text = {
-            Column {
+            // Scrollable, because this dialog grew past a screenful: a Material dialog constrains
+            // its own height and clips what does not fit, so the last control on the list is one
+            // large system font away from being unreachable with nothing on screen to say so.
+            Column(Modifier.verticalScroll(rememberScrollState())) {
                OutlinedTextField(
                    value = name,
                    onValueChange = { name = it },
@@ -212,6 +308,70 @@ fun SessionSettingsDialog(
                    )
                }
                Spacer(Modifier.height(8.dp))
+                Row(
+                    verticalAlignment = Alignment.CenterVertically,
+                    modifier = Modifier.fillMaxWidth(),
+                ) {
+                    Text("Resume after a usage limit", modifier = Modifier.weight(1f))
+                    if (autoResume == null && resumeError == null) {
+                        CircularProgressIndicator(
+                            modifier = Modifier.width(16.dp).height(16.dp),
+                            strokeWidth = 2.dp,
+                        )
+                        Spacer(Modifier.width(8.dp))
+                    }
+                    Switch(
+                        checked = autoResume == true,
+                        onCheckedChange = { setAutoResume(it, resumeMessage) },
+                        enabled = autoResume != null,
+                    )
+                }
+                // Disabled rather than hidden while the switch is off: a field that comes and goes
+                // makes its own presence the signal, and a visible one teaches what the switch will
+                // do. Committed on the keyboard's Done rather than on every keystroke, so typing a
+                // sentence is one request instead of one per letter.
+                OutlinedTextField(
+                    value = resumeMessage,
+                    onValueChange = { resumeMessage = it },
+                    label = { Text("Message to send") },
+                    // What an empty field means, in the field: the server's own word rather than a
+                    // session poked with nothing to read.
+                    placeholder = { Text(DEFAULT_RESUME_MESSAGE) },
+                    singleLine = true,
+                    enabled = autoResume == true,
+                    modifier = Modifier.fillMaxWidth(),
+                    keyboardOptions = KeyboardOptions(imeAction = ImeAction.Done),
+                    keyboardActions =
+                        KeyboardActions(onDone = { setAutoResume(true, resumeMessage) }),
+                )
+                // What it does and what it costs, in the order it happens. The last sentence is the
+                // one that matters: the time below is when the server will *ask*, not a promise
+                // about when the session speaks.
+                Text(
+                    "When this session stops because the account is out of quota, the server " +
+                        "checks the limit and sends this message once it has lifted. It checks " +
+                        "again if the limit is still on.",
+                    style = MaterialTheme.typography.bodySmall,
+                    color = MaterialTheme.colorScheme.onSurfaceVariant,
+                )
+                // Only where something is actually waiting. Absent is not a state worth a row: a
+                // session that has not hit a limit has nothing scheduled, which the reader can see
+                // from the switch.
+                resumeAt?.let { at ->
+                    Text(
+                        "Waiting now -- next check ${formatCheckTime(at)}.",
+                        style = MaterialTheme.typography.bodySmall,
+                        color = MaterialTheme.colorScheme.onSurfaceVariant,
+                    )
+                }
+                resumeError?.let {
+                    Text(
+                        it,
+                        color = MaterialTheme.colorScheme.error,
+                        style = MaterialTheme.typography.bodySmall,
+                    )
+                }
+                Spacer(Modifier.height(8.dp))
                Row(
                    verticalAlignment = Alignment.CenterVertically,
                    modifier = Modifier.fillMaxWidth(),
@@ -257,6 +417,44 @@ fun SessionSettingsDialog(
                        style = MaterialTheme.typography.bodySmall,
                    )
                }
+                // Left out rather than disabled, the one place this dialog does that: a disabled
+                // control teaches what the thing can do, and a llama session cannot do this at all
+                // -- the row would be teaching something false about it.
+                if (takesEffort) {
+                    Spacer(Modifier.height(8.dp))
+                    Row(
+                        verticalAlignment = Alignment.CenterVertically,
+                        modifier = Modifier.fillMaxWidth(),
+                    ) {
+                        Text("Thinking", modifier = Modifier.weight(1f))
+                        PickerButton(
+                            current = level ?: DEFAULT_EFFORT,
+                            // The level the CLI picks for itself is in the list as well as in the
+                            // button, so leaving a level is not a one-way trip -- the same
+                            // correction the model picker carries.
+                            options = listOf(DEFAULT_EFFORT) + EFFORT_LEVELS,
+                            onPick = { chosen ->
+                                setEffort(chosen.takeIf { it != DEFAULT_EFFORT })
+                            },
+                        )
+                    }
+                    // What it costs, said where it is about to be pressed, like Move above: the
+                    // CLI reads the level when it launches and has no control request for
+                    // changing one.
+                    Text(
+                        "Changing this stops the session's process. It starts again with the " +
+                            "next message, or with Start.",
+                        style = MaterialTheme.typography.bodySmall,
+                        color = MaterialTheme.colorScheme.onSurfaceVariant,
+                    )
+                    effortError?.let {
+                        Text(
+                            it,
+                            color = MaterialTheme.colorScheme.error,
+                            style = MaterialTheme.typography.bodySmall,
+                        )
+                    }
+                }
                Spacer(Modifier.height(8.dp))
                Row(
                    verticalAlignment = Alignment.CenterVertically,
@@ -317,6 +515,21 @@ fun SessionSettingsDialog(
                    Text("Render timings", modifier = Modifier.weight(1f))
                    TextButton(onClick = onCopyRenderReport) { Text("Copy") }
                }
+                // Bench-build only: see [onRunBenchmark]. Named exactly "Run benchmark" because
+                // ui-trace and the emulator smoke run find it by that label, the same way every
+                // other control here is found -- see AGENTS.md's "Driving the UI".
+                onRunBenchmark?.let { run ->
+                    Spacer(Modifier.height(8.dp))
+                    Row(
+                        verticalAlignment = Alignment.CenterVertically,
+                        modifier = Modifier.fillMaxWidth(),
+                    ) {
+                        Glyph(SPEED_GLYPH, colour = MaterialTheme.colorScheme.onSurface)
+                        Spacer(Modifier.width(8.dp))
+                        Text("P0 benchmark", modifier = Modifier.weight(1f))
+                        TextButton(onClick = run) { Text("Run benchmark") }
+                    }
+                }
            }
        },
        // Disabled rather than absent while there is nothing to save: a button that comes and goes
@@ -329,3 +542,21 @@ fun SessionSettingsDialog(
        dismissButton = { TextButton(onClick = onDismiss) { Text("Close") } },
    )
 }
+
+/**
+ * When the server will next look, as a local time.
+ *
+ * A time rather than a countdown, for the reason the transcript's own limit row gives: this screen
+ * reads the figure once, and a span drawn from a value nothing refreshes goes stale while somebody
+ * is looking at it.
+ */
+private fun formatCheckTime(epochSeconds: Double): String =
+    try {
+        DateTimeFormatter.ofLocalizedTime(FormatStyle.SHORT)
+            .withZone(ZoneId.systemDefault())
+            .format(Instant.ofEpochSecond(epochSeconds.toLong()))
+    } catch (_: Exception) {
+        // A time that cannot be read is not a time to show: the sentence above still says a check
+        // is coming, which is the part the reader can act on.
+        "soon"
+    }
@@ -70,13 +70,22 @@ class UsageFeed(
    /** Ask the backend again now. The dialog's refresh button; the poll does it on its own. */
    val refresh: () -> Unit,
 ) {
-    /** What [setup]'s own limits came back as. See [usageFor] for why the states are these. */
-    fun forSetup(setup: String): SessionUsage =
-        when (val state = snapshots) {
+    /**
+     * What meters [session], and what that meter came back as. See [usageFor] for the states.
+     *
+     * A session rather than a machine, because a machine is not what is metered: one machine runs
+     * the Claude CLI and an echo session side by side, and only the first of them spends anything.
+     */
+    fun forSession(session: SessionSummary): SessionUsage {
+        // Settled without asking anybody: a session nothing meters has nothing to check, and
+        // "checking" is what the fetch's own states would say about it for as long as one is out.
+        val provider = session.usageProvider ?: return SessionUsage.NotMetered
+        return when (val state = snapshots) {
            is LoadState.Loading -> SessionUsage.Waiting
            is LoadState.Error -> SessionUsage.Unavailable(state.message)
-            is LoadState.Loaded -> usageFor(state.value, setup)
+            is LoadState.Loaded -> usageFor(state.value, session.setup, provider)
        }
+    }
 }

 /**
@@ -158,9 +167,15 @@ fun SessionUsageBar(usage: SessionUsage, modifier: Modifier = Modifier) {
        }
    }

-    // Nothing at all for a machine that meters nothing: a row saying "unknown" there would report a
-    // problem about a setup somebody chose, on every screen, forever.
-    if (usage is SessionUsage.NotMetered) {
+    // Nothing at all for a session that meters nothing: a row saying "unknown" there would report
+    // a problem about a setup somebody chose, on every screen, forever.
+    //
+    // And nothing while the first fetch is out, which is a different silence. A request in flight
+    // is not a state to report -- and the session that meters nothing is exactly the one this
+    // cannot yet tell apart, so "5-hour usage: checking" appeared under an echo session for half a
+    // second and was then taken away. A row that has to be withdrawn is worse than one that
+    // arrives late.
+    if (usage is SessionUsage.NotMetered || usage is SessionUsage.Waiting) {
        return
    }

@@ -171,9 +186,10 @@ fun SessionUsageBar(usage: SessionUsage, modifier: Modifier = Modifier) {
        // Words, not a colour and not an empty bar: every one of these is a different kind of
        // answer from "this much is used", and only words carry a difference in kind.
        when (val state = usage) {
-            SessionUsage.NotMetered -> Unit
+            // Both handled above, before the row exists at all.
+            SessionUsage.NotMetered,
+            SessionUsage.Waiting -> Unit
            is SessionUsage.Unavailable -> UsageNote("5-hour usage unknown -- ${state.why}")
-            SessionUsage.Waiting -> UsageNote("5-hour usage: checking")
            is SessionUsage.Known -> {
                val window = state.windows.firstOrNull { it.kind == "session" }
                if (window == null) {
@@ -234,16 +250,22 @@ private fun fiveHourLabel(window: UsageWindow, now: OffsetDateTime): String {
 }

 /**
- * One machine's snapshot, out of every machine's.
+ * One meter's snapshot, out of every machine's: [setup]'s row for [provider].
+ *
+ * Both halves are needed to pick it. A machine can hold more than one meter -- the Claude CLI's
+ * account and, while a test has one set, an echo session's invented one -- and a snapshot is one
+ * service on one machine.
 *
 * Every way of having *failed* to get numbers is [SessionUsage.Unavailable] with the reason in it.
 * None of them may look like zero, and none may look like [SessionUsage.NotMetered], which is the
 * machine having no quota rather than the question going unanswered.
 */
-fun usageFor(snapshots: List<UsageSnapshot>, setup: String): SessionUsage {
-    // No snapshot at all means the backend never asked, which it only does for a machine with
-    // nothing metered on it. That is a different answer from having asked and failed.
-    val mine = snapshots.firstOrNull { it.setup == setup } ?: return SessionUsage.NotMetered
+fun usageFor(snapshots: List<UsageSnapshot>, setup: String, provider: String): SessionUsage {
+    // No snapshot at all means the backend never asked, which it only does where there is nothing
+    // to ask about. That is a different answer from having asked and failed.
+    val mine =
+        snapshots.firstOrNull { it.setup == setup && it.provider == provider }
+            ?: return SessionUsage.NotMetered
    if (mine.state != "ok") {
        return SessionUsage.Unavailable(mine.detail ?: mine.state)
    }
@@ -236,6 +236,7 @@ private fun AddSetupDialog(
    var address by remember { mutableStateOf("") }
    var identity by remember { mutableStateOf("") }
    var attachmentsDir by remember { mutableStateOf("") }
+    var modelsDir by remember { mutableStateOf("") }
    var tested by remember { mutableStateOf<String?>(null) }
    var testing by remember { mutableStateOf(false) }

@@ -250,6 +251,7 @@ private fun AddSetupDialog(
                    port = typedPort,
                    identityFile = identity.trim().ifEmpty { null },
                    attachmentsDir = attachmentsDir.trim().ifEmpty { null },
+                    modelsDir = modelsDir.trim().ifEmpty { null },
                )
            }

@@ -293,6 +295,14 @@ private fun AddSetupDialog(
                    label = { Text("Folder for attached files (optional)") },
                    singleLine = true,
                )
+                // Where that machine's GGUFs are, for a llama.cpp session on it. Blank means
+                // the same place this backend keeps its own downloads, read on that machine.
+                OutlinedTextField(
+                    value = modelsDir,
+                    onValueChange = { modelsDir = it },
+                    label = { Text("Folder for models (optional)") },
+                    singleLine = true,
+                )
                tested?.let {
                    Spacer(Modifier.height(8.dp))
                    Text(it, style = MaterialTheme.typography.bodySmall)
@@ -61,18 +61,31 @@ fun SpawnScreen(
    // "auto" rather than "manual": on a phone every ask is a round trip to a question card, and
    // answering "allow Bash?" dozens of times per task is what this app exists to avoid.
    var permissionMode by remember { mutableStateOf("auto") }
+    // Null until the server has been asked, and null again if it answers "no level chosen" -- the
+    // two are told apart by [defaultsAsked], because a picker that shows a level before the answer
+    // arrives is one you can spawn at without having chosen it.
+    var effort by remember { mutableStateOf<String?>(null) }
+    var defaultsAsked by remember { mutableStateOf(false) }
    var busy by remember { mutableStateOf(false) }
    // Only the spawn's own failure. The fetch's lives in `options`: this one leaves a filled-in
    // form worth keeping, and that one leaves nothing to fill in.
    var spawnError by remember { mutableStateOf<String?>(null) }
-    // Downloaded models, for a llama provider to choose between. Kept separate from the setups: a
-    // Claude session needs none, so failing to list them must not stop the screen rendering.
+    // The models on the *chosen machine*, for a llama provider to choose between. Kept separate
+    // from the setups: a Claude session needs none, so failing to list them must not stop the
+    // screen rendering. Refetched when the machine changes, because a model is a file on one
+    // machine -- see [fetchSetupModels].
    var models by remember { mutableStateOf<List<LocalModel>>(emptyList()) }
    var modelKey by remember { mutableStateOf<String?>(null) }
    var contextSize by remember { mutableStateOf("") }
    var temperature by remember { mutableStateOf("") }

    LaunchedEffect(Unit) {
+        // Separate from the setups fetch below and deliberately not fatal: failing to learn the
+        // default must leave a screen you can still spawn from, so the picker stays on "default"
+        // and says so rather than the whole form refusing to draw.
+        runCatching { withContext(Dispatchers.IO) { fetchDefaultEffort(settings) } }
+            .onSuccess { effort = it }
+        defaultsAsked = true
        options =
            try {
                val fetched = withContext(Dispatchers.IO) { fetchSetups(settings) }
@@ -83,9 +96,6 @@ fun SpawnScreen(
            } catch (e: ApiException) {
                LoadState.failed(e)
            }
-        models =
-            runCatching { withContext(Dispatchers.IO) { fetchModels(settings).local } }
-                .getOrDefault(emptyList())
    }

    Column(Modifier.fillMaxSize().verticalScroll(rememberScrollState()).padding(16.dp)) {
@@ -115,6 +125,17 @@ fun SpawnScreen(
                is LoadState.Loaded -> state.value
            }
        val setup = setups.firstOrNull { it.name == setupName }
+        // Whichever machine is chosen now, asked again when that changes. The old machine's list
+        // is dropped first rather than left on screen: a file name from another machine looks
+        // exactly like one from this one.
+        LaunchedEffect(setup?.id) {
+            models = emptyList()
+            modelKey = null
+            val id = setup?.id ?: return@LaunchedEffect
+            models =
+                runCatching { withContext(Dispatchers.IO) { fetchSetupModels(settings, id) } }
+                    .getOrDefault(emptyList())
+        }
        val current = setup?.providers?.firstOrNull { it.name == providerName }
        // Only the Claude CLI has models, a working directory and permission modes; keying the
        // extra fields on the kind rather than the provider name keeps a second Claude provider
@@ -175,12 +196,13 @@ fun SpawnScreen(
        )

        if (isLlama) {
-            // A llama session names one of the models this backend has downloaded, so the choice is
-            // that list rather than free text -- a name that is not on disk is a session that
-            // cannot start.
+            // A llama session names one of the models on the machine it will run on, so the
+            // choice is that list rather than free text -- a name that is not on that machine's
+            // disk is a session that cannot start.
            if (models.isEmpty()) {
                Text(
-                    "No models downloaded yet. Get one from the Models screen first.",
+                    "No models on ${setup?.name ?: "this machine"}. The Models screen downloads " +
+                        "to the backend; another machine needs the file put there itself.",
                    style = MaterialTheme.typography.bodyMedium,
                    color = MaterialTheme.colorScheme.onSurfaceVariant,
                )
@@ -251,6 +273,19 @@ fun SpawnScreen(
                selected = permissionMode,
                onSelect = { permissionMode = it },
            )
+            Spacer(Modifier.height(16.dp))
+
+            // Says what it does to *later* spawns as well, because it does: the level chosen here
+            // is stored as the default, which is the whole way that default is set. A picker that
+            // quietly changed a global would be the same control with the fact left out.
+            ChipGroup(
+                label = "Thinking (kept as the default for new sessions)",
+                options = listOf(DEFAULT_EFFORT) + EFFORT_LEVELS,
+                // The CLI's own default is a level in the list, so this cannot be a one-way trip.
+                // Disabled-looking until the server has answered, for the reason above.
+                selected = if (defaultsAsked) effort ?: DEFAULT_EFFORT else null,
+                onSelect = { chosen -> effort = chosen.takeIf { it != DEFAULT_EFFORT } },
+            )
        }
        Spacer(Modifier.height(24.dp))

@@ -268,6 +303,13 @@ fun SpawnScreen(
                    try {
                        val spawned =
                            withContext(Dispatchers.IO) {
+                                // Stored before the spawn and not after it: choosing a level is
+                                // an intent about new sessions in general, so a spawn that then
+                                // fails must not also lose the choice. Non-fatal for the same
+                                // reason the fetch above is -- the session is what was asked for.
+                                if (isClaude) {
+                                    runCatching { setDefaultEffort(settings, effort) }
+                                }
                                spawnSession(
                                    settings,
                                    // The id, not the label: labels are editable and the server
@@ -280,6 +322,7 @@ fun SpawnScreen(
                                        if (isLlama) modelKey else model.trim().takeIf { isClaude },
                                    cwd = cwd.trim().takeIf { isClaude },
                                    permissionMode = permissionMode.takeIf { isClaude },
+                                    effort = effort.takeIf { isClaude },
                                    // Sent only when set, so blank means "whatever llama.cpp does
                                    // by default" rather than a zero.
                                    params =
@@ -0,0 +1,27 @@
+package com.example.aiapp
+
+/**
+ * Where one transcript lives: a session's own, or one of its subagents'.
+ *
+ * The single mechanism [fetchTranscript], [EventStream], [TranscriptSource] and
+ * [TranscriptCache.session] all take, rather than each growing its own branch between a session and
+ * a subagent -- see SUBAGENTS.md's "Phone" and "Wire shape". A caller that has only a session id
+ * builds one with the one-argument constructor; a subagent's screen supplies both ids.
+ */
+data class TranscriptAddress(val sessionId: String, val subagentId: String? = null) {
+    /** The URL segment naming this transcript, before `/transcript` or `/events`. */
+    val urlPath: String
+        get() =
+            if (subagentId == null) "sessions/$sessionId"
+            else "sessions/$sessionId/subagents/$subagentId"
+
+    /**
+     * Where this transcript's cache lives on the phone, relative to the cache root.
+     *
+     * A subagent's nests under its session's directory rather than sitting beside it, so deleting a
+     * session's cache directory takes its subagents' with it -- the same one-way door the server's
+     * own storage describes.
+     */
+    val cachePath: String
+        get() = if (subagentId == null) sessionId else "$sessionId/subagents/$subagentId"
+}
@@ -35,8 +35,15 @@ class TranscriptCache(
    private val root: File,
    private val warn: (String) -> Unit = { Log.w("ai-app", it) },
 ) {
-    /** The cache for one session, whether or not anything has been stored for it yet. */
-    fun session(id: String): SessionCache = SessionCache(File(root, id), warn)
+    /**
+     * The cache for one transcript, whether or not anything has been stored for it yet.
+     *
+     * A subagent's [TranscriptAddress.cachePath] nests it under its session's directory, so
+     * deleting the session (below) takes its subagents' caches with it -- there is no separate
+     * purge for one.
+     */
+    fun session(address: TranscriptAddress): SessionCache =
+        SessionCache(File(root, address.cachePath), warn)

    /**
     * Deletes every session directory not in [ids], called after a successful list fetch. The path
@@ -175,6 +175,18 @@ sealed class TranscriptItem {
        val preTokens: Long?,
        val postTokens: Long?,
    ) : TranscriptItem()
+
+    /**
+     * The account ran out of quota, so the turn stopped here.
+     *
+     * A divider rather than an error: nothing failed, and what a reader scrolling back needs from
+     * it is the same thing a clear or a compaction gives them -- why the conversation stops at this
+     * line.
+     *
+     * [resetsAt] is epoch seconds and null where the session was told nothing, which is a state the
+     * row has words for rather than a time it invents.
+     */
+    data class LimitNote(override val seq: Long, val resetsAt: Double?) : TranscriptItem()
 }

 /**
@@ -458,6 +470,7 @@ fun foldEvent(items: List<TranscriptItem>, entry: SeqEvent): List<TranscriptItem
            } else {
                items + TranscriptItem.ImageItem(entry.seq, event.ref)
            }
+        is SessionEvent.LimitReached -> items + TranscriptItem.LimitNote(entry.seq, event.resetsAt)
        is SessionEvent.Cleared -> items + TranscriptItem.ClearedNote(entry.seq)
        is SessionEvent.Compacted ->
            items + TranscriptItem.CompactedNote(entry.seq, event.preTokens, event.postTokens)
@@ -18,7 +18,7 @@ import java.util.concurrent.atomic.AtomicReference
 */
 class TranscriptSource(
    private val settings: ServerSettings,
-    private val sessionId: String,
+    private val address: TranscriptAddress,
    val cache: SessionCache,
 ) {
    private val stream = AtomicReference<EventStream?>(null)
@@ -65,7 +65,7 @@ class TranscriptSource(
        val tail = cache.tail() ?: return false
        // `before = seq + 1` is the newest event with seq <= the cursor, which is the event *at*
        // the cursor when the server still has one there.
-        val answer = fetchTranscript(settings, sessionId, before = tail.seq + 1, limit = 1)
+        val answer = fetchTranscript(settings, address, before = tail.seq + 1, limit = 1)
        val matches =
            answer.size == 1 &&
                try {
@@ -83,7 +83,7 @@ class TranscriptSource(
     */
    suspend fun fetchOpening(): List<SeqEvent> {
        DebugStats.count("transcript page from server")
-        val page = fetchTranscript(settings, sessionId, limit = OPENING_WINDOW)
+        val page = fetchTranscript(settings, address, limit = OPENING_WINDOW)
        page.forEach { (line, entry) -> cache.append(line, entry.seq) }
        cache.flush()
        return page.map { it.second }
@@ -108,7 +108,7 @@ class TranscriptSource(
        val page =
            fetchTranscript(
                settings,
-                sessionId,
+                address,
                before = before,
                limit = limit,
                coalesce = coalesce,
@@ -131,7 +131,7 @@ class TranscriptSource(
     * well lose.
     */
    fun follow(after: Long, onOpen: () -> Unit, onReset: () -> Unit, onEvent: (SeqEvent) -> Unit) {
-        val opened = EventStream(settings, sessionId)
+        val opened = EventStream(settings, address)
        stream.getAndSet(opened)?.close()
        try {
            opened.run(after, onOpen, onReset) { raw, entry ->
@@ -0,0 +1,6 @@
+<?xml version="1.0" encoding="utf-8"?>
+<resources>
+    <!-- Overridden by the `bench` build type's resValue (build.gradle.kts) to "AI Sessions bench",
+         so the two are never mistaken for each other in the launcher or in Settings. -->
+    <string name="app_name">AI Sessions</string>
+</resources>
@@ -0,0 +1,32 @@
+package com.example.aiapp
+
+import java.time.ZoneId
+import kotlin.test.Test
+import kotlin.test.assertEquals
+import kotlin.test.assertTrue
+
+/**
+ * What the transcript says where a session ran out of quota.
+ *
+ * The pair worth a test is the one that reads the same when it goes wrong: a reset time that
+ * arrived and one that never did. The second must not turn into a plausible-looking time, because a
+ * reader has no way of telling an invented one from a reported one.
+ */
+class LimitRowTest {
+    private val utc = ZoneId.of("UTC")
+
+    @Test
+    fun `a reported reset time is shown as a time`() {
+        // 2026-09-05T12:00:00Z. Asserted as a prefix and the clock reading rather than as the
+        // whole string: the platform's own short-time format is what this asks for, and it
+        // differs by JDK and locale down to which space character separates the meridiem.
+        val summary = limitSummary(1_788_609_600.0, utc)
+        assertTrue(summary.startsWith("Usage limit reached • resets "), summary)
+        assertTrue(summary.contains("12:00"), summary)
+    }
+
+    @Test
+    fun `a limit with no reset time says only what is known`() {
+        assertEquals("Usage limit reached", limitSummary(null, utc))
+    }
+}
@@ -23,7 +23,7 @@ class TranscriptCacheTest {

    private fun cache() = TranscriptCache(File(temp, "v1/host_8443")) { said += it }

-    private fun session(id: String = "s") = cache().session(id)
+    private fun session(id: String = "s") = cache().session(TranscriptAddress(id))

    private fun line(seq: Long, type: String = "toolStart") =
        """{"seq":$seq,"ts":1.5,"type":"$type","id":"x"}"""
@@ -0,0 +1,28 @@
+# The P0 benchmark fixture
+
+`transcript.jsonl` is a synthetic transcript in the app's own event model (the JSON lines
+`GET /sessions/{id}/transcript` returns; see `Events.kt`'s `parseSeqEvent` and
+`server/src/session/driver.rs`) -- never a real one. It is what both the Compose `bench` build
+and iris's bench build open with no server, so the two apps draw exactly the same content and a
+frame-time comparison is measuring the renderer rather than the data.
+
+Generated by `./generate.py` (Python stdlib only, seeded -- `SEED = 20260905` -- so re-running it
+reproduces the same file byte for byte). It writes into `assets/` -- a separate directory from this
+script and README, because the Compose `bench` build type points its own asset source set straight
+at `assets/` (`app/androidApp/build.gradle.kts`'s `sourceSets { getByName("bench") }`), and a Python
+script and a markdown file have no business inside an APK:
+
+- `transcript.jsonl` -- 3,601 events. The first 3,200 (`BACKLOG_COUNT`) are the scrolled-back
+  history the benchmark opens with: user turns, tool calls with kilobyte-scale input/output,
+  assistant replies built from headings, bold/italic/inline code, a link, fenced code blocks that
+  rotate through rust/kotlin/python/sh/json/toml, a markdown table, two embedded images, and
+  periodic `usageDelta`/`compacted` events. The remaining 400 (`STREAM_COUNT`) are not part of the
+  opening window -- both bench harnesses replay them at a fixed rate (20/s) through the same live
+  fold path a real SSE reply arrives on, which is P0's "streaming phase."
+- `bench1.png`, `bench2.png` -- tiny (8x8) flat-colour PNGs, base64-free on disk but served the
+  same way a real attachment is (`GET /sessions/{id}/files/{name}`), referenced by the two
+  `"type":"image"` events in the transcript.
+
+Regenerate after changing the shape (a new event type, a different backlog/stream split) with
+`./generate.py`, and commit the result -- it is checked in rather than generated at build time so
+both apps' bench builds embed the identical bytes without needing this script at build time.
@@ -0,0 +1,186 @@
+#!/usr/bin/env python3
+"""Generates transcript.jsonl -- the synthetic fixture P0's benchmark opens in both apps.
+
+Deterministic (fixed seed), so a Compose bench APK and an iris bench APK draw byte-identical
+content: the point of the fixture is a like-for-like comparison, not a realistic one.
+
+Never a real transcript -- see AGENTS.md's ui-sandbox.sh, which this borrows its vocabulary
+style from (headings, code fences, a table, a link) rather than reusing its Claude-Code JSONL
+shape. This file's shape is the *app's own event model* instead: one JSON object per line,
+matching what GET /sessions/{id}/transcript returns and what Events.kt's parseSeqEvent reads
+(server/src/session/driver.rs is the source of truth for the field names).
+
+    ./generate.py                      writes transcript.jsonl and bench1.png/bench2.png here
+
+BACKLOG_COUNT events (seq 1..BACKLOG_COUNT) are the scrolled-back history the benchmark opens
+with. A further STREAM_COUNT events (seq BACKLOG_COUNT+1..) are not part of the opening window;
+both bench harnesses replay them at a fixed rate as the "streaming reply" phase, appended through
+the same live path a real SSE reply arrives on. Keeping both halves in one file means one
+generator and one seed to keep in sync, rather than two fixtures that can drift apart.
+"""
+import base64
+import json
+import random
+import struct
+import zlib
+from pathlib import Path
+
+SEED = 20260905
+BACKLOG_COUNT = 3200
+STREAM_COUNT = 400
+HERE = Path(__file__).resolve().parent / "assets"
+
+random.seed(SEED)
+
+LANGUAGES = ["rust", "kotlin", "python", "sh", "json", "toml"]
+
+CODE_SNIPPETS = {
+    "rust": '''fn fold_event(items: Vec<Item>, seq: u64) -> Vec<Item> {
+    // a comment worth keeping: this is the fold the app's own screen runs
+    let mut out = items;
+    out.push(Item::new(seq));
+    out
+}''',
+    "kotlin": '''fun foldEvent(items: List<TranscriptItem>, entry: SeqEvent): List<TranscriptItem> {
+    // mirrors the server's own event model, one item per line
+    return items + TranscriptItem.from(entry)
+}''',
+    "python": '''def render_report(frames, cpu_ms, rss_kb):
+    # printed for a human to paste back, so every number carries its unit
+    return f"{frames} frames, {cpu_ms}ms cpu, {rss_kb}kb peak rss"''',
+    "sh": '''#!/bin/sh
+# scripted scroll loop, the shape transcript-bench.sh drives on a phone
+for i in $(seq 1 24); do
+    ui-trace record --do "swipe 540 700 540 1600 200"
+done''',
+    "json": '{"seq": 1, "type": "status", "state": "running"}',
+    "toml": '''[package]
+name = "bench-fixture"
+version = "0.1.0"''',
+}
+
+HEADINGS = [
+    "## Plan",
+    "## What changed",
+    "## Why this approach",
+    "### Open questions",
+    "## Results",
+]
+
+WORDS = (
+    "session render report frame budget scroll transcript fold event cache "
+    "cursor probe stream backlog swipe fixture bench compose iris widget layout "
+    "measure place draw tool call token context window anchor"
+).split()
+
+
+def paragraph(n=24):
+    words = [random.choice(WORDS) for _ in range(n)]
+    words[0] = words[0].capitalize()
+    text = " ".join(words) + "."
+    # Sprinkle markdown inline spans so the syntax highlighter/markdown parser sees a real mix.
+    text = text.replace(" fold ", " **fold** ", 1)
+    text = text.replace(" cursor ", " *cursor* ", 1)
+    text = text.replace(" cache ", " `cache` ", 1)
+    if "bench" in text:
+        text = text.replace(
+            " bench ", " [bench](https://example.com/bench) ", 1
+        )
+    return text
+
+
+def make_png(rgb, size=8):
+    """A tiny, valid PNG -- flat colour, no external dependency."""
+
+    def chunk(tag, data):
+        c = tag + data
+        return struct.pack(">I", len(data)) + c + struct.pack(">I", zlib.crc32(c))
+
+    sig = b"\x89PNG\r\n\x1a\n"
+    ihdr = struct.pack(">IIBBBBB", size, size, 8, 2, 0, 0, 0)
+    raw = b""
+    for _ in range(size):
+        raw += b"\x00" + bytes(rgb) * size
+    idat = zlib.compress(raw)
+    return sig + chunk(b"IHDR", ihdr) + chunk(b"IDAT", idat) + chunk(b"IEND", b"")
+
+
+def main():
+    HERE.mkdir(exist_ok=True)
+    lines = []
+    seq = 1
+    ts = 1_788_000_000.0
+
+    def emit(type_, **fields):
+        nonlocal seq, ts
+        obj = {"seq": seq, "ts": round(ts, 3), "type": type_}
+        obj.update(fields)
+        lines.append(json.dumps(obj, separators=(",", ":")))
+        seq += 1
+        ts += random.uniform(0.05, 2.0)
+
+    emit("status", state="running")
+    emit("settings", model="bench-model", permissionMode="auto")
+
+    image_refs = []
+    turn = 0
+    while seq <= BACKLOG_COUNT:
+        turn += 1
+        emit("userMessage", text=f"Turn {turn}: {paragraph(12)}", id=None, attachments=[])
+
+        # A tool call with kilobyte-scale input/output every few turns.
+        if turn % 3 == 0:
+            tool_id = f"tool-{turn}"
+            big_input = json.dumps({"path": f"/repo/file_{turn}.rs", "content": paragraph(400)})
+            emit("toolStart", id=tool_id, tool="Edit", input=big_input)
+            big_output = "\n".join(paragraph(60) for _ in range(20))
+            emit("toolUpdate", id=tool_id, output=big_output[: len(big_output) // 2])
+            emit("toolEnd", id=tool_id, output=big_output)
+
+        # A reply: a heading, prose, a fenced block in a rotating language, a table, then deltas.
+        emit("assistantText", delta=f"{random.choice(HEADINGS)}\n\n")
+        emit("assistantText", delta=paragraph(30) + "\n\n")
+        lang = LANGUAGES[turn % len(LANGUAGES)]
+        emit("assistantText", delta=f"```{lang}\n{CODE_SNIPPETS[lang]}\n```\n\n")
+        if turn % 5 == 0:
+            emit(
+                "assistantText",
+                delta="| column | value |\n|---|---|\n| a | " + paragraph(3) + " |\n\n",
+            )
+        # A run of small deltas -- the shape a live reply actually streams in.
+        for _ in range(random.randint(3, 8)):
+            emit("assistantText", delta=paragraph(6) + " ")
+
+        # A couple of images, base64 PNGs, the way a real transcript embeds a screenshot.
+        if turn in (10, 40):
+            ref = f"bench{len(image_refs) + 1}.png"
+            image_refs.append(ref)
+            emit("image", ref=ref, about=None)
+
+        emit("usageDelta", tokens=random.randint(200, 4000), context=random.randint(2000, 180000))
+
+        if turn % 15 == 0:
+            emit(
+                "compacted",
+                preTokens=180000,
+                postTokens=20000,
+                trigger="auto",
+            )
+
+    # The streaming-phase tail: one long reply, built entirely from text deltas, the shape a
+    # bench harness replays at a fixed events/sec through the live fold path.
+    emit("userMessage", text="One more, streamed live for the benchmark's timing phase.", id=None, attachments=[])
+    while seq <= BACKLOG_COUNT + STREAM_COUNT:
+        emit("assistantText", delta=paragraph(5) + " ")
+    emit("status", state="idle")
+
+    (HERE / "transcript.jsonl").write_text("\n".join(lines) + "\n")
+
+    (HERE / "bench1.png").write_bytes(make_png((220, 90, 90)))
+    (HERE / "bench2.png").write_bytes(make_png((90, 150, 220)))
+
+    print(f"wrote {len(lines)} events ({BACKLOG_COUNT} backlog + {STREAM_COUNT} stream) to transcript.jsonl")
+
+
+if __name__ == "__main__":
+    main()
@@ -4,6 +4,11 @@
 #   ./build-apk.sh            the release build, signed (what the phone runs)
 #   ./build-apk.sh debug      the debug build, for reproducing something the
 #                             emulator scripts would build anyway
+#   ./build-apk.sh bench      P0's benchmark build (own app id, "AI Sessions
+#                             bench" label, opens straight onto the fixture
+#                             session -- see docs/RUST.md's P0 box and
+#                             app/bench-fixture/README.md). Signed the same
+#                             as release; never touches the CA it pins.
 #
 # Dev Updater's `.dev-updater.ron` at the checkout root spells these out as
 # build modes, one command line each; it passes nothing else, so the word
@@ -25,8 +30,9 @@ VARIANT=${1:-release}
 case "$VARIANT" in
 release) TASK=assembleRelease ;;
 debug) TASK=assembleDebug ;;
+bench) TASK=assembleBench ;;
 *)
-    echo "build-apk.sh: unknown variant '$VARIANT' (release, debug)" >&2
+    echo "build-apk.sh: unknown variant '$VARIANT' (release, debug, bench)" >&2
    exit 2
    ;;
 esac
@@ -81,7 +87,7 @@ fi
 # uninstalling it first: the signatures differ, and Android refuses to
 # update across them.
 KEYSTORE="${AI_APP_KEYSTORE:-${XDG_CONFIG_HOME:-$HOME/.config}/ai-app/release.jks}"
-if [ "$VARIANT" = release ] && [ ! -f "$KEYSTORE" ]; then
+if { [ "$VARIANT" = release ] || [ "$VARIANT" = bench ]; } && [ ! -f "$KEYSTORE" ]; then
    KEYTOOL="${JAVA_HOME:+$JAVA_HOME/bin/keytool}"
    KEYTOOL="${KEYTOOL:-keytool}"
    if ! command -v "$KEYTOOL" >/dev/null 2>&1; then
@@ -97,7 +103,7 @@ if [ "$VARIANT" = release ] && [ ! -f "$KEYSTORE" ]; then
        -keyalg RSA -keysize 2048 -validity 10000 \
        -storepass "$PASSWORD" -keypass "$PASSWORD" -dname "CN=ai-app" >/dev/null 2>&1)
 fi
-if [ "$VARIANT" = release ]; then
+if [ "$VARIANT" = release ] || [ "$VARIANT" = bench ]; then
    AI_APP_KEYSTORE="$KEYSTORE"
    AI_APP_KEYSTORE_PASSWORD=$(cat "$KEYSTORE.password")
    export AI_APP_KEYSTORE AI_APP_KEYSTORE_PASSWORD
@@ -0,0 +1,34 @@
+#!/bin/sh
+# RUST.md's I5 "Where iris's frame time goes" pass (2026-09-05). The same
+# 24-swipe/6-cycle loop as transcript-bench.sh's, extracted for iris's own
+# demo app -- transcript-bench.sh itself is Compose-specific (opens by
+# session title through the Compose app's own UI) and cannot be called
+# directly against dev.iris.android.demo.
+#
+# MUST be run from inside this checkout (not /tmp): ui-trace/adb pick which
+# emulator to target from the current directory's basename (the
+# per-checkout-AVD rule), and a previous pass lost two attempts to a `cd`
+# into /tmp that made this resolve to a nonexistent "tmp" checkout.
+set -eu
+cd "$(dirname "$0")"
+. ./android-env.sh >/dev/null 2>&1
+
+cycles=${1:-6}
+
+ui-trace record -d 3000 --do "tap 'Reset frame report'" -o /tmp/iris-bench-reset.txt >/dev/null
+adb logcat -c
+
+DO=""
+i=0
+while [ "$i" -lt "$cycles" ]; do
+    DO="$DO --do 'swipe 540 700 540 1600 200' --do 'wait 500'"
+    DO="$DO --do 'swipe 540 700 540 1600 200' --do 'wait 500'"
+    DO="$DO --do 'swipe 540 1600 540 700 200' --do 'wait 500'"
+    DO="$DO --do 'swipe 540 1600 540 700 200' --do 'wait 500'"
+    i=$((i + 1))
+done
+eval ui-trace record -d $((cycles * 16000 + 20000)) $DO -o /tmp/iris-bench-scroll.txt >/dev/null
+
+ui-trace record -d 3000 --do "tap 'Frame report'" -o /tmp/iris-bench-report.txt >/dev/null
+sleep 1
+adb logcat -d -s iris-android-app:I | grep "iris frame report:"
@@ -7,7 +7,7 @@ edition = "2024"
 # with `server/` via `event-model`), the REST + SSE clients for its HTTP
 # surface (see `server/src/routes.rs`'s module doc for the table), the
 # transcript fold and cache, the markdown block model, the syntax
-# highlighter and the ANSI parser. See CLIENT_CORE.md at the repo root for
+# highlighter and the ANSI parser. See `docs/CLIENT_CORE.md` for
 # what this holds today, what it does not yet, and how it corresponds to
 # the Kotlin it replaces.
 #
@@ -78,14 +78,14 @@ fn percent_decode(s: &str) -> String {
    let mut out = Vec::with_capacity(bytes.len());
    let mut i = 0;
    while i < bytes.len() {
-        if bytes[i] == b'%' && i + 2 < bytes.len() {
-            if let Ok(byte) =
+        if bytes[i] == b'%'
+            && i + 2 < bytes.len()
+            && let Ok(byte) =
                u8::from_str_radix(std::str::from_utf8(&bytes[i + 1..i + 3]).unwrap_or(""), 16)
-            {
-                out.push(byte);
-                i += 3;
-                continue;
-            }
+        {
+            out.push(byte);
+            i += 3;
+            continue;
        }
        out.push(bytes[i]);
        i += 1;
@@ -1,5 +1,5 @@
 //! The app's pure logic, shared between the server and any Rust client --
-//! see `CLIENT_CORE.md` at the repo root for what lives here and what does
+//! see `docs/CLIENT_CORE.md` for what lives here and what does
 //! not yet.

 pub mod ansi;
@@ -1,7 +1,7 @@
 //! This phone's copy of the transcripts it has already been sent, so
 //! reopening a session does not download it again. Ported from
-//! `app/.../TranscriptCache.kt`; see `TRANSCRIPT_CACHE.md` at the repo root
-//! for the design and `CLIENT_CORE.md` for how this file corresponds to it.
+//! `app/.../TranscriptCache.kt`; see `docs/TRANSCRIPT_CACHE.md`
+//! for the design and `docs/CLIENT_CORE.md` for how this file corresponds to it.
 //!
 //! What is stored is the server's own JSON for one event per line, in
 //! transcript order. Reading the cache means running the same [`seq_of`]
@@ -606,6 +606,37 @@ pub fn group_tool_runs(items: &[TranscriptItem]) -> Vec<TranscriptRow> {
    rows
 }

+/// Folds a page of raw transcript lines (`ApiClient::fetch_transcript_page`'s
+/// `Vec<Value>`) into the flat item list this module works over. A line
+/// this build can't parse fails the whole page rather than being skipped --
+/// CODE_RULES's "an enumeration must be able to say 'it broke'" -- since
+/// silently dropping one event could hide, say, a user message that then
+/// looks like it was never sent. Moved here from `desktop-app`'s `app.rs`
+/// (RUST.md's E4) when the Android transcript client (I5) needed the same
+/// fold: "write the logic once" applies to any caller embedding
+/// `transcript-ui` against a live server, not just the first one.
+pub fn fold_page(values: &[serde_json::Value]) -> Result<Vec<TranscriptItem>, String> {
+    let mut items = Vec::new();
+    for value in values {
+        let event: SeqEvent = serde_json::from_value(value.clone()).map_err(|e| {
+            format!("the server sent a transcript line this build couldn't parse: {e}")
+        })?;
+        items = fold_event(&items, &event);
+    }
+    Ok(items)
+}
+
+/// The wire `seq` a raw transcript line carries -- the live-stream resume
+/// cursor after loading a page must be this, not a folded item's `seq()`.
+/// A folded `AssistantMsg` keeps the seq of the *first* delta it
+/// accumulated (`fold_event`'s own doc), so resuming from that seq would
+/// re-deliver every delta already folded into it, duplicating the tail of
+/// a reply that was mid-stream when the page was fetched -- found via a
+/// real screenshot in E4 (RUST.md), where the assistant's line doubled.
+pub fn raw_seq(value: &serde_json::Value) -> Option<u64> {
+    value.get("seq")?.as_u64()
+}
+
 #[cfg(test)]
 mod tests {
    use super::*;
@@ -824,4 +855,88 @@ mod tests {
            other => panic!("expected a QuestionCard, got {other:?}"),
        }
    }
+
+    fn line(seq: u64, json: serde_json::Value) -> serde_json::Value {
+        let mut obj = json;
+        obj["seq"] = serde_json::json!(seq);
+        obj["ts"] = serde_json::json!(1.0);
+        obj
+    }
+
+    /// The regression for a bug a real `run-headless.sh` screenshot found
+    /// in `desktop-app` (E4, RUST.md): resuming the live stream from the
+    /// last *item's* seq re-delivers the deltas already folded into a
+    /// still-open assistant message, doubling its tail. `raw_seq` of the
+    /// last wire line must be the true high-water mark instead, which for a
+    /// run of deltas is higher than every item's own `seq()`.
+    #[test]
+    fn the_resume_cursor_is_the_last_wire_seq_not_the_last_items_seq() {
+        let values = vec![
+            line(1, serde_json::json!({"type": "userMessage", "text": "hi"})),
+            line(
+                2,
+                serde_json::json!({"type": "assistantText", "delta": "a"}),
+            ),
+            line(
+                3,
+                serde_json::json!({"type": "assistantText", "delta": "b"}),
+            ),
+            line(
+                4,
+                serde_json::json!({"type": "assistantText", "delta": "c"}),
+            ),
+        ];
+        let after = raw_seq(values.last().unwrap()).unwrap();
+        assert_eq!(after, 4);
+
+        let items = fold_page(&values).unwrap();
+        let assistant_seq = items
+            .iter()
+            .find(|i| matches!(i, TranscriptItem::AssistantMsg { .. }))
+            .unwrap()
+            .seq();
+        assert_eq!(assistant_seq, 2);
+        assert_ne!(
+            after, assistant_seq,
+            "the fixed bug: these must differ here"
+        );
+    }
+
+    #[test]
+    fn a_page_folds_into_one_settled_assistant_message() {
+        let values = vec![
+            line(1, serde_json::json!({"type": "userMessage", "text": "hi"})),
+            line(
+                2,
+                serde_json::json!({"type": "assistantText", "delta": "hel"}),
+            ),
+            line(
+                3,
+                serde_json::json!({"type": "assistantText", "delta": "lo"}),
+            ),
+        ];
+        let items = fold_page(&values).unwrap();
+        assert_eq!(
+            items,
+            vec![
+                TranscriptItem::UserMsg {
+                    seq: 1,
+                    text: "hi".to_string(),
+                    attachments: Vec::new(),
+                },
+                TranscriptItem::AssistantMsg {
+                    seq: 2,
+                    text: "hello".to_string(),
+                    settled: false,
+                },
+            ]
+        );
+    }
+
+    #[test]
+    fn an_unparseable_line_fails_the_whole_page() {
+        let values = vec![serde_json::json!({"seq": 1, "ts": 1.0, "type": "not-a-real-type"})];
+        let err = fold_page(&values).unwrap_err();
+        assert!(err.contains("couldn't parse"));
+    }
 }
@@ -0,0 +1,287 @@
+# Decisions taken for Iris to review
+
+Short list of design choices made by the design agent without asking, so
+they can be judged and reversed later. Detail lives in RUST.md (and IRIS.md
+for iris API changes); this file is only the summary. Newest first. Items
+marked **DEFERRED** are ones the agent chose not to decide alone.
+
+## 2026-09-05
+
+- **iris no longer asks every device for compute-shader limits it never
+  uses.** `adapter.request_device` (both `iris/src/android/render.rs` and
+  `iris/src/default/render.rs`) used `Limits::default()` plus an override
+  for `max_buffer_size`, and `Limits::default()` unconditionally requests
+  desktop-tier compute limits (`max_compute_workgroups_per_dimension:
+  65535`, per `wgpu_types`) even though nothing in `iris`/`iris-core`
+  creates a `ComputePipeline` or writes a `@compute` shader stage —
+  confirmed by grepping the whole tree, not assumed. That crashed
+  `request_device` outright on the Android emulator's software GL path
+  (`EMU_GPU=software`, `--features force-gles`): SwiftShader's GL reports
+  itself as OpenGL ES 3.0, which has no compute shaders at all, so the
+  adapter's real limit is 0 against the unconditional request for 65535 —
+  `RUST.md`'s "Software mode ... crashes for a third, different reason,"
+  2026-09-05, earlier today. The same would happen on any real
+  GLES-3.0-only Android device, not just the emulator. Fixed by a new
+  `iris_core::device_limits()` (`iris/core/src/render/mod.rs`), shared by
+  both platform backends so the two requests cannot drift, that zeros the
+  six `max_compute_*` fields explicitly rather than switching to a
+  downlevel `Limits` preset — `Limits::downlevel_webgl2_defaults()` was
+  considered and rejected: it also zeros
+  `max_storage_buffers_per_shader_stage`, and `shader.wgsl`'s vertex stage
+  reads four `var<storage>` buffers (rects, glyphs, masks, move_offsets),
+  so that preset would trade the compute crash for a bind-group-layout
+  one on the same downlevel hardware this is meant to support. No
+  capability check or fallback path was needed since nothing is being
+  disabled — the request is simply narrowed to what the pipeline actually
+  uses. `rigs/gpu-probe`'s own mirrored limits (it is deliberately its own
+  crate, not a workspace member, so it cannot call `device_limits()`
+  directly) were updated to match, and confirm `IRIS DEVICE: ok` against
+  this VM's own Vulkan and GL adapters. **Not verified this pass**: the
+  specific SwiftShader-ES-3.0 crash this fixes, on-device — the
+  `EMU_GPU=software` cold boot this needs would have force-restarted this
+  checkout's emulator while another session was actively running its own
+  app on it (`com.example.aiapp` had window focus at the time), so it was
+  left for a pass when the emulator is free rather than disrupting that
+  session. Everything reachable without the emulator is clean: `cargo
+  fmt`/`clippy --workspace --all-targets`/`test --workspace`, `cargo ndk
+  build`/`clippy` for `iris-android-app` with `force-gles`, and
+  `gpu-probe` against this VM's own Vulkan and GL(ES 3.2, which still has
+  compute and so would not have reproduced the crash even before this
+  fix — not a substitute for the real ES-3.0 test).
+
+- **P0's Compose half is built and smoke-tested on the emulator** — the
+  `bench` build type, the shared `app/bench-fixture/` transcript, and an
+  in-process fake backend (`BenchFixture.kt`/`BenchNetwork.kt`) that
+  answers `TranscriptSource`/`EventStream` from an in-memory event log
+  instead of a real server, so the fold and paging under test are the real
+  ones. Full account, the smoke run's report, and what is deliberately
+  left (the iris half, the real on-phone runs) are in RUST.md's P0 box.
+  Not a decision to review so much as the gate itself now being runnable —
+  flagged here because it is the first half of something Iris explicitly
+  asked to see before P1.
+
+- **P0's iris half is also built and smoke-tested on the emulator,
+  2026-09-05.** A new `bench` Cargo feature on `iris-android-app`, on top
+  of `transcript-screen`: the same checked-in fixture (`include_str!`, no
+  asset pipeline needed), the same 24-swipe scroll loop animated through
+  `List::scroll` and the same 400-event/20s streaming phase through
+  `fold_event`, "Run benchmark"/"Copy report" as named accessible
+  controls, and the same three added report fields (process CPU time,
+  peak RSS, battery current) via direct JNI calls
+  (`bench_jni.rs::PlatformHandle`) since `android_view` has no
+  `BatteryManager`/`ClipboardManager` wrapper of its own. One small public
+  API addition to get there: `AndroidAppState::platform_ready` (`IRIS.md`),
+  a default-no-op lifecycle hook handing an implementor a `JavaVM` +
+  `GlobalRef` it can call Java through from any thread. Packaged with a
+  new `release` build type on `iris-android-app`'s own Gradle project
+  (there was previously only `debug`), signed with the same key
+  `app/build-apk.sh` generates. Smoke run and the full report are in
+  RUST.md's P0 box; not attempted this pass: the real on-phone runs and
+  Iris's pass/fail call, which is the actual gate.
+
+- **The intermittent touch-scroll dropout is root-caused and fixed: a
+  missed `ACTION_DOWN` hit-test, not the previously-suspected coalesced
+  first `ACTION_MOVE`.** Diagnosed by temporary logcat tracing of every
+  touch event, `DragArbiter` state transition and `Selection::drag`
+  dispatch (removed once confirmed), reproduced on this checkout's own
+  emulator against a real sandbox session. The trace showed the actual
+  mechanism: a gesture's `ACTION_DOWN` lands wherever the finger actually
+  is, which is not guaranteed to fall inside the same row-local sensor
+  region a later `ACTION_MOVE` in the same gesture lands in (a row's own
+  padding/gap, or its non-selectable sender-name header, is
+  pointer-transparent to `iris::sense::CursorSense`). When that happens,
+  the widget that ends up handling the gesture never saw `PressStart`, so
+  `DragArbiter` sits in `Idle` — which answers every subsequent frame with
+  `Undecided` and has no way to tell "no press is happening" from "a press
+  is happening but I missed its start," so it never recovers on its own
+  for the rest of that gesture. One real trace showed exactly this: touch
+  `Down`/`Move`/`Up` all delivered correctly, but zero `PressStart`
+  reaching the arbiter, `state=Idle` unchanged from first frame to last.
+  Fixed at the call site that has the context to recover
+  (`iris::transcript_ui::selection::Selection::drag`,
+  `iris/transcript-ui/src/selection.rs`): a new `DragArbiter::is_idle()`
+  (`iris/src/sense.rs`) lets it notice a `Pressing` frame arriving with the
+  arbiter still `Idle` — which can only mean a missed `PressStart`, since a
+  `Pressing` sense requires the button to genuinely be down — and start the
+  press there instead of where it was missed. Three new unit tests in
+  `sense.rs`'s `drag_arbiter_tests` and one in `transcript-ui`'s
+  `selection::tests` (the latter fails on the code before this fix).
+  Commit follows. Not the same failure the earlier pass's `DECISIONS.md`
+  DEFERRED item speculated about (a coalesced first `ACTION_MOVE` skipping
+  slop detection) — that hypothesis is now ruled out; the arbiter's own
+  slop/long-press logic was never wrong. RUST.md's I5 box,
+  "Touch-scroll dropout root-caused, 2026-09-05" has the full trace.
+- **P0, a phone benchmark gate before any porting, asked for by Iris
+  2026-09-05**: "before P1 I'd like to see benchmarks & also maybe stress
+  test on my own phone ... If it doesn't match compose reasonably well then
+  I don't think I'd wanna continue." Design (RUST.md's P0 box has the
+  detail): the same embedded synthetic fixture in both apps with no server
+  needed; the same scripted scroll loop then a streaming phase, run
+  programmatically since the phone has no usable system tracing and no
+  agent can drive it; the same report from both (frames, janky %, p50/p90/
+  p99, process CPU time, peak RSS, battery current where readable) with a
+  copy button; the iris app under its own id and the Compose one as a new
+  `bench` build type with an id suffix, so neither replaces her production
+  install; two arm64 APKs plus instructions delivered under `~/host/bench/`.
+  The gate is hers: iris within a reasonable margin of Compose release on
+  p50, p99 and CPU time, no crashes, no visible stutter. If it fails, the
+  port stops.
+- **The rest of the port is one UI crate, `iris/app-ui`, grown out of
+  `iris/transcript-ui` rather than started beside it.** It holds a
+  `Screen` enum plus a back stack — the Rust equivalent of `AppRoot.kt`'s
+  `when` — and `iris/desktop-app`/`iris/android-app` become thin entry
+  points over it. Chosen over a fresh crate because `transcript-ui`
+  already has the right generic shape (`Rsc: HasEvents` +
+  `Rsc::State: FocusHost`) and the `client-core`/`event-model` path
+  dependencies every later screen needs, so growing it in place is the
+  smaller diff. Platform-only code (notification service, share target,
+  QR scanner, Keystore token, deep-link enrolment) stays in the E3/E5
+  Java shell (`android-shell/` + `app/shellApp`) rather than moving into
+  this crate, since none of it is a screen. The Android APK is built by
+  `cargo xtask apk` (E5), merging the app-ui cdylib into the E3 shell so
+  there is one app rather than a demo shell plus a service shell.
+  `app/androidApp` (the Compose app) stays untouched and is the baseline
+  every step is measured against, until parity is reached (P7 decides
+  the switch, and is itself a load-bearing decision left to Iris). Order
+  is by risk to the daily-use path: session screen first (P1, where
+  every hard behaviour already lives), then the shell merge and a real
+  phone install (P2), then root tabs (P3), the explorer (P4),
+  settings/enrolment (P5), desktop parity (P6), and the cutover itself
+  (P7). Full plan: RUST.md's "The port, in order (decided 2026-09-05)".
+- **iris gets its own measured frame report, rather than waiting on a
+  `dumpsys`/`gfxinfo` answer that cannot see a `SurfaceView`'s GPU-drawn
+  frames.** `iris_core::FrameReport` (`iris/core/src/render/frame_report.rs`)
+  times each frame's wall clock from the same point `render()`'s redraw
+  starts to just after `queue.submit` + `present()` — the span Compose's
+  own render report and `gfxinfo` both count — into a fixed 4096-entry
+  ring (no allocation per frame; `report()` is the only place that
+  allocates, and only on a button tap). The report gives total frames,
+  janky % over the same 16.7ms budget `gfxinfo` uses, P50/P90/P99 and the
+  worst, plus a reset. Exposed the way the Compose app's copy-button
+  report already is: two named controls ("Frame report", "Reset frame
+  report") on the transcript screen, tappable by accessibility name via
+  `ui-trace`, logging under this crate's fixed `android_logger` tag
+  (`iris-android-app`) so a script can grep `"iris frame report"` the way
+  `transcript-bench.sh` greps `"ai-app render report"`. The report's own
+  `Display` line says plainly that it measures up to the `present()` call
+  returning, not GPU/compositor completion — wgpu's `present()` is not
+  fenced against either, so presenting that span as "time to reach the
+  screen" would be a measured-looking number that is actually inferred,
+  which the standing UI rule forbids.
+- **`ui-trace` gains a hold-then-drag gesture, additive, in
+  `emulator-tools`.** Neither of its two existing actions can produce
+  "hold stationary for `LONG_PRESS`, then move without lifting" — `tap`
+  has no hold and `swipe X1 Y1 X2 Y2 MS` interpolates motion across its
+  whole duration from t=0. A new action presses, waits, then moves to a
+  second point and releases as one continuous touch (raw
+  `sendevent`/`MotionEvent` injection, extending whatever mechanism the
+  existing `swipe` already uses), so `DragArbiter`'s pan-vs-select rule
+  (`iris/src/sense.rs`, already covered by 8 unit tests against a
+  synthetic clock) can finally be driven on a real device instead of only
+  in a test harness.
+- **Touch drag on a transcript row follows Android's own rule**: a vertical
+  drag pans the list immediately; a stationary press held 500 ms starts a
+  text selection which further dragging extends; a horizontal drag while
+  something is already selected extends that selection without the wait.
+  One `DragArbiter` per list decides it (`iris/src/sense.rs`). Chosen over a
+  "text layer always wins" or "list always wins" rule because either loses
+  one of the two gestures a reader expects.
+- **E4's desktop shape is a new `iris/desktop-app` crate**: a winit window
+  holding `transcript-ui`'s screen beside a session list, talking to a real
+  `ai-server` through `client-core`. It enrols by pasting the same
+  `aiapp://enroll?…` link a phone scans (`client-core::config::EnrolledServer`)
+  and keeps it owner-only under `$XDG_CONFIG_HOME/ai-app-desktop/`. The
+  pinned CA is a path given on the command line, not baked in. Chosen so
+  the phone and desktop share one enrolment format and no second one is
+  invented.
+- **I5's Android integration extends `iris-android-app` (I2's shell)
+  behind a Cargo feature (`transcript-screen`), rather than a third
+  shell crate.** That project already has the Gradle module, the
+  `IrisView`/`MainActivity` Java, and the JNI registration; the only
+  thing a second screen needs on top is a different `AndroidAppState`,
+  the same axis `tabs_ui::build`/`transcript_ui::build` already vary
+  along on the winit side. `tabs-screen`/`transcript-screen` are
+  mutually exclusive and each pulls in only its own deps, so the plain
+  tabs build (I2/I4) is untouched.
+- **Order of remaining work, updated 2026-09-05**: the two in-flight
+  pieces and I5's Android integration are all done; next is giving iris
+  its own frame-timing report so item 3 below can be decided by a number.
+- **DECIDED by Iris, 2026-09-05: iris is the app's framework; Masonry was
+  the calibration.** Her words: "I think iris definitely makes more sense
+  based on the limitations we've found." The limitations: Masonry has no
+  touch scroll on Android (E2), no per-span rich text and no cross-row
+  selection on the pinned commit (E2), and its keyboard bridge is a TODO
+  (E1); iris carries the same screen under the Compose baseline on the
+  host GPU (p50 15.0 ms against Compose's 20.0 ms, RUST.md's I5 box). What
+  follows: the E-steps are closed as calibration, and the port proceeds
+  on iris — screens, the shell (E3/E5), and `client-core` underneath.
+  The item below is kept as the record of what she decided from.
+- **Was DEFERRED — whether to commit to iris over Masonry for `ai-app`.**
+  Updated 2026-09-05 with the clean comparison the recommendation wanted:
+  same sandbox session content, same emulator, `EMU_GPU=software`, one
+  session. Headline numbers (RUST.md's I5 box, "Clean scroll comparison,
+  2026-09-05," has the full table and every caveat):
+
+  | app | build | frames | janky % | p50 | p90 | p99 | worst |
+  |---|---|---|---|---|---|---|---|
+  | Compose (in-app report) | debug | 1102 | 99.0% late | 33.8ms | 50.6ms | 79.5ms | -- |
+  | Compose (`dumpsys gfxinfo`) | debug | 1499 | 21.15% (95.66% legacy) | 32ms | 48ms | 150ms (p99) | -- |
+  | iris (`FrameReport`) | **release** | 299 | 94.65% | 79.1ms | 98.6ms | 117.8ms | 212.6ms |
+  | iris (`FrameReport`, repeat) | **release** | 233 | 94.42% | 109.3ms | 130.8ms | 147.1ms | 150.5ms |
+
+  **Not a clean apples-to-apples reading, stated plainly rather than
+  smoothed over**: iris had to be built **release** (debug `SIGSEGV`s on
+  this emulator's Vulkan loader, I4's finding) against Compose's mandated
+  **debug** build, so this asymmetry likely *understates* iris's gap
+  rather than the reverse; the three frame-time sources measure different
+  things (Compose's own phase accounting vs. Android's HWUI deadline-miss
+  definition vs. iris's redraw-start-to-present window, the last of which
+  `dumpsys gfxinfo` cannot see at all for iris's `SurfaceView`); and both
+  figures are emulator numbers under software rasterisation, which
+  Compose's *own* in-app report shows already costs 20-34ms/frame in
+  `swap`+`gpu` alone under this GPU mode, so a same-mode iris number well
+  above 16.7ms was expected going in for either app. A second pair under
+  `-gpu host` was not taken this pass. The earlier session's suspected
+  intermittent touch-delivery dropout was **not reproduced** this pass —
+  the zero-frame results this time traced to this pass's own script bug
+  (a `cd` that changed which emulator `ui-trace` targeted), not the
+  emulator; a CPU-load rise during the gesture was observed by a sampler
+  running throughout, but did not correlate with any failure, so the
+  original candidate is neither confirmed nor ruled out.
+  The choice in front of Iris, updated: decide now on the
+  structural-plus-functional case already made (iris works end-to-end
+  where Masonry's scroll gesture doesn't exist at all on Android) plus
+  this table — reading the two build profiles and three jank definitions
+  with the caveats above rather than as a single number — or ask for a
+  same-profile, same-GPU-mode rerun first. RUST.md's I5 box has the full
+  account.
+
+  **Updated 2026-09-05, the `-gpu host` pair taken.** Real GPU rendering
+  (`force-gles` -- the default Vulkan backend has no adapter at all under
+  plain host-GPU boot, confirmed by the exact `wgpu` error) reverses the
+  software-mode shape:
+
+  | app | build | GPU mode | frames | janky % | p50 | p90 | p99 | worst | cpu p50 | gpu-wait p50 |
+  |---|---|---|---|---|---|---|---|---|---|---|
+  | Compose (in-app report) | debug | host (virgl) | 1268 | 96.4% late | 20.0ms | 28.4ms | 37.7ms | -- | -- | -- |
+  | iris (`FrameReport`) | **release**, `force-gles` | host (virgl) | 62 | 41.94% | 15.0ms | 21.8ms | 37.1ms | 37.1ms | 0.2ms | 12.9ms |
+
+  Under real GPU rendering iris's median frame is *faster* than
+  Compose's, not the 2-3x-slower shape the software-mode table shows. A
+  new split inside `FrameReport` (redraw-to-submit vs. submit-to-present,
+  commit `e2a1fad`) says why: iris's own CPU work per frame is a median
+  0.2ms -- almost the entire frame is time spent handing the frame to the
+  driver, not in iris's layout/text/primitive code. This is consistent
+  with the earlier software-mode gap being mostly SwiftShader's CPU
+  rasterisation cost rather than an iris-specific slowness, but is not
+  proof of it: a same-mode software `force-gles` run to isolate the
+  backend crashed for an unrelated reason (SwiftShader's GL path reports
+  itself as OpenGL ES 3.0, which has no compute shaders, and iris's device
+  request assumes them unconditionally) — real scope to fix, not done
+  here — and the two apps' frame populations still differ in kind the same
+  way the software-mode caveats describe. A real intermittent touch-
+  scroll dropout was also reproduced this pass (six consecutive swipes
+  produced zero redraws while taps kept working; an identical retry then
+  succeeded) and is not explained. RUST.md's I5 box, "Where iris's frame
+  time goes, 2026-09-05, the `-gpu host` pass," has the full account. The
+  iris-vs-Masonry choice itself is still Iris's to make.
@@ -8,6 +8,127 @@ capability that moved. Small and trivial changes do not go here.
 An entry gives the date, what changed, why, and a short before/after where
 it helps judge the change without the session that made it. Newest first.

+## 2026-09-05: `AndroidAppState::platform_ready` (RUST.md's P0 box, iris half)
+
+Added a second, optional lifecycle method to `iris::android::AndroidAppState`
+(`iris/src/android/view.rs`), called once from `new_peer` right after `new`:
+
+```rust
+fn platform_ready(&mut self, rsc: &mut AndroidRsc<Self>, vm: JavaVM, view: GlobalRef) {}
+```
+
+Default does nothing, so every existing implementor (`Client`,
+`TranscriptClient`) is unaffected. It exists for a caller that needs to call
+into Java itself beyond what a `RequestRedraw` handle already covers --
+P0's bench build (`iris-android-app`'s new `bench` feature,
+`bench_client.rs`/`bench_jni.rs`) uses it to hold a `JavaVM` + `GlobalRef`
+to the view so its "Copy report" control and once-a-second battery sampler
+can call `BatteryManager`/`ClipboardManager` through the view's own
+`Context`, from a background tokio task as well as the UI thread. `new`
+itself was not extended with these two parameters: most implementors need
+nothing here, and `new`'s job is building the widget tree, not holding a
+platform handle. `vm`/`view` are independent handles from the ones
+`new_peer` keeps for its own `RequestRedraw` (a fresh `get_java_vm`/
+`new_global_ref` each), so storing them has no effect on that mechanism.
+
+## 2026-09-05 (later still): `iris_core::device_limits()`, and iris no longer requests compute-shader limits
+
+New public function, `iris_core::device_limits() -> wgpu::Limits`. Why:
+`adapter.request_device`'s `required_limits` was `Limits::default()` plus
+a `max_buffer_size` override in both platform backends, and
+`Limits::default()` requests desktop-tier compute-shader limits
+unconditionally (`max_compute_workgroups_per_dimension: 65535`) even
+though nothing in `iris`/`iris-core` uses a `ComputePipeline` — that
+crashed device creation outright on a downlevel GL adapter reporting
+OpenGL ES 3.0 (no compute shaders at all: the Android emulator's
+`EMU_GPU=software` path, and any real GLES-3.0-only Android device).
+`device_limits()` is what both `android::render::AndroidRenderer::new`
+and `default::render::UiRenderer::new` now build their `required_limits`
+from, so the request cannot drift between the two backends.
+
+Before: `Limits { max_buffer_size: 1 << 30, ..Default::default() }`
+inlined in each backend. After: `iris_core::device_limits()`, which is
+the same thing with the six `max_compute_*` fields additionally zeroed.
+A caller building its own `DeviceDescriptor` outside these two backends
+(there are none today, but a third platform backend would want this)
+should call `device_limits()` rather than reaching for
+`Limits::default()` directly, unless it genuinely adds a compute pass —
+in which case it wants the specific compute limits that pass needs, not
+the desktop-tier default for everything.
+
+## 2026-09-05 (later the same day): `iris_core::FrameReport` (RUST.md's I5 box)
+
+New public type, `iris_core::FrameReport` (re-exported from `iris_core`'s
+`render` module alongside `FrameStats` and `JANK_THRESHOLD`). Why: `dumpsys
+gfxinfo` cannot see a `SurfaceView`'s own GPU-drawn frames at all, so a
+`wgpu`-rendered iris screen had no way to ask "was this smooth" the way
+Compose's own in-app render report already can -- item 3 of RUST.md's
+recommendation was stuck on a one-sided number for exactly this reason.
+
+`FrameReport::record(elapsed: Duration)` is called once per frame (wired
+into `android/view.rs`'s `render()`, wrapping the same span from redraw
+start to after `queue.submit`+`present()` that Compose's report and
+`gfxinfo` both count) and writes into a fixed 4096-entry ring -- no
+allocation on the hot path. `FrameReport::report() -> Option<FrameStats>`
+gives total frames, janky % (over `JANK_THRESHOLD`, the same 16.7ms 60Hz
+budget `gfxinfo` uses), P50/P90/P99 and the worst; `None` if nothing has
+been recorded since the last `reset()`, not a zeroed report that would
+read as a real measurement. `FrameStats`'s `Display` line says plainly
+that it measures up to `present()` being called, not GPU/compositor
+completion, since wgpu's `present()` isn't fenced against either.
+
+`AndroidUiState` gained a `pub frame_report: FrameReport` field --
+anything with `HasAndroidUiState` can now read or reset it. Before this,
+there was no way to ask iris's own render path how long a frame took at
+all, on any backend.
+
+Before/after, for a caller that already has `ui_state: &AndroidUiState`:
+
+```rust
+// before: no such question could be asked
+// after:
+match ui_state.frame_report.report() {
+    Some(stats) => log::info!("iris frame report: {stats}"),
+    None => log::info!("iris frame report: no frames recorded yet"),
+}
+ui_state.frame_report.reset(); // via android_state_mut()
+```
+
+`iris-android-app`'s transcript screen exposes this as two named,
+tappable controls ("Frame report", "Reset frame report") rather than
+requiring a caller to wire its own UI -- see `transcript_client.rs`'s
+`frame_report_controls`.
+
+## 2026-09-05: `Tasks::redraw_handle` (RUST.md's I5 Android integration)
+
+New public method on `iris::task::Tasks`, `redraw_handle(&self) ->
+Arc<dyn RequestRedraw>`. Why: a caller running its own long-lived loop
+*inside* one spawned task (a live SSE follow, the Android transcript
+client's `select_session`) has no other way to ask for a frame after each
+`TaskCtx::update` -- `Tasks::spawn`'s own wrapper only requests one, after
+the whole async closure finishes, which fits a single request-then-update
+but not a stream that needs to be seen redrawing after *each* event. This
+is the same gap `iris/desktop-app`'s module doc names for why it uses
+winit's `Proxy<AppEvent>` instead of `Tasks` -- android-view has no
+`Proxy`, so this is what closes it there.
+
+**A real bug this uncovered, not a hypothetical**: calling the returned
+handle's `request_redraw()` from the background thread crashed the process
+(`SIGABRT`, `Result::unwrap() on an Err value: JavaException`) the first
+time an Android transcript fetch called it a second time. `android/render.rs`'s
+`AndroidRedrawHandle` was already attaching the calling thread to the JVM
+correctly, but its `request_redraw` called `View::post_frame_callback`,
+whose Java side calls `Choreographer.getInstance()` -- which throws unless
+the *calling* thread already has a `Looper`, and a tokio worker thread,
+even freshly JNI-attached, has none. Fixed by routing through
+`View::post_delayed(0)` instead (Android's own thread-safe "queue work onto
+this View's UI thread" primitive, needing no caller-side `Looper`), landing
+on a new `IrisViewPeer::delayed_callback` override that drains tasks and
+renders -- same body as `do_frame`, on the UI thread where
+`post_frame_callback` is safe again. Any future caller of `redraw_handle()`
+from a background thread gets this for free; nothing about the fix is
+specific to the transcript screen.
+
 ## 2026-09-05: `transcript_ui::build_tree` (RUST.md's E4)

 `transcript_ui::build` claimed the whole window (`ui_state.set_root(tree)`)
@@ -73,6 +194,24 @@ on a row's own rendered text now pans the list correctly instead of
 always starting a selection. 8 new unit tests in `iris/src/sense.rs`'s
 `drag_arbiter_tests` module.

+### 2026-09-05, later: `DragArbiter::is_idle()`, recovering a missed `press_start`
+
+Follow-up to the above, from a real touch-scroll dropout: a gesture's
+`ACTION_DOWN` can land on a caller's own dead space (a row's padding, a
+gap, a header with no handler) that never calls `press_start`, so the
+first frame the arbiter actually sees is a `Pressing`-shaped `update`
+with no matching start. Before this, `update`'s `Idle` arm had no way to
+tell that apart from "nothing is happening" and answered `Undecided`
+forever for the rest of that gesture. `is_idle(&self) -> bool` lets a
+caller notice the gap and recover: if `is_idle()` is true on a frame the
+caller knows a press is genuinely down (its own `Pressing`/equivalent
+sense fired), call `press_start` right there instead of assuming one
+already happened. `transcript-ui`'s `Selection::drag` is the reference
+caller — one new match arm, checked before the ordinary `update`-only
+case. Any other `DragArbiter` caller with the same "one sensor per
+sub-region, no fallback for dead space" shape has the same gap and wants
+the same recovery.
+
 ## 2026-09-05: `SpanStyle`, per-range text styling (RUST.md's I5)

 A `TextBuffer` used to have exactly one style (`TextAttrs`: colour, size,
@@ -275,3 +414,26 @@ TEXTURES.md's "Recommended shape" and "Implemented, 2026-09-04".
  `wgpu::BindGroup` and draw call instead of a slot in the shared array —
  invisible from the widget API, visible only in `UiRenderNode`'s internals
  and in `iris`'s device requirements.
+
+## 2026-09-05: `FrameReport` splits each frame at `queue.submit`
+
+`FrameStats` gains two fields, and `FrameReport` gains a second recording
+method, to answer "is a slow frame iris's own CPU work or the driver/GPU"
+with a number instead of a guess (RUST.md's I5 box).
+
+- **`FrameReport::record_split(total, submit_to_present)`** is a second way
+  to record a frame, alongside the existing `record(total)` (unchanged,
+  and still what a caller with no split should use — it now reads as
+  `cpu_p50 == total`, `gpu_wait_p50 == 0`, rather than fabricating a
+  number for a half it never measured).
+- **`FrameStats` gains `cpu_p50` and `gpu_wait_p50`**: medians of
+  redraw-start-to-submit and submit-to-after-`present()` respectively,
+  independent of each other and of the existing `p50`/`p90`/`p99`/`worst`
+  (which are unchanged, and still over the whole frame). The Android
+  renderer's `draw()` now returns the `submit_to_present` `Duration` it
+  measured, which `android::view::render()` passes to `record_split`.
+- **Caveat carried in both doc comments**: `submit_to_present` is not
+  fenced against the GPU actually finishing — it is "how long the CPU was
+  blocked handing the frame to the driver," not a confirmed GPU-completion
+  time. Enough to separate "iris is slow building the frame" from "iris is
+  slow handing it off," not enough to claim an exact GPU budget.
@@ -7,6 +7,29 @@ order and what "done" looks like. Tick and date them in place.

 ## Fix

+- [x] **`request_device` asked for compute-shader limits it never uses
+  (2026-09-05).** `Limits::default()` (both `iris/src/android/render.rs`
+  and `iris/src/default/render.rs`) requests desktop-tier compute limits
+  unconditionally, even though nothing in `iris`/`iris-core` creates a
+  `ComputePipeline` or writes a `@compute` shader stage — confirmed by
+  grepping the whole tree, not assumed. That crashed device creation
+  outright on the Android emulator's software GL path (`EMU_GPU=software`,
+  `--features force-gles`): SwiftShader's GL reports itself as OpenGL ES
+  3.0, which has no compute shaders, so the adapter's real limit is 0
+  against the unconditional request for 65535 — the same would happen on
+  any real GLES-3.0-only Android device. Fixed by a new, shared
+  `iris_core::device_limits()` (`iris/core/src/render/mod.rs`) that zeros
+  exactly the six `max_compute_*` fields rather than switching to a
+  downlevel `Limits` preset — `downlevel_webgl2_defaults()` also zeros
+  `max_storage_buffers_per_shader_stage`, which `shader.wgsl`'s vertex
+  stage needs (four `var<storage>` buffers), so that preset would trade
+  this crash for a bind-group-layout one on the same hardware.
+  `rigs/gpu-probe`'s own hand-mirrored `Limits` (it is deliberately its
+  own crate, not able to call `device_limits()` directly) was updated to
+  match. See `DECISIONS.md` and RUST.md's I5 box for the account,
+  including what could not be re-verified on-device this pass (the
+  emulator was in concurrent use by another session).
+
 - [x] **Input does not fall through by input type (2026-09-04).**
  `SensorUi::run_sensors` (`src/default/sense.rs`) used to set "consumed,
  stop checking lower layers" from mere hover — a widget registered for
@@ -82,6 +105,18 @@ order and what "done" looks like. Tick and date them in place.
  were exactly the same root cause measured two different ways. Frame 2
  now reports 0 (see the numbers above); not a separate fix.

+- [ ] **A read-only text display has no widget of its own — P0's bench
+  report area is a `TextEdit` standing in for one (2026-09-05).** The only
+  way to get selectable text on screen today is `.editable(...)` plus
+  `.attr::<Selectable>(())` (`Selectable` is only implemented for
+  `TextEdit`, `iris/src/attr.rs`), which also makes the field focusable —
+  tapping the bench report opens the soft keyboard over text nothing lets
+  you type into. Harmless for a bench-only debug screen (not fixed this
+  pass), but a real "selectable, not editable" text primitive would
+  remove the keyboard side effect and is worth having before another
+  screen wants the same thing (P1's own transcript rows already read
+  their content from a `TextEdit` for the same reason).
+
 ## Build

 - [x] **Benchmarks**, not unit tests, run on demand (2026-09-05; a
@@ -176,14 +211,62 @@ order and what "done" looks like. Tick and date them in place.
  silently dropped. See RUST.md's I5 box for the full account of what
  *was* built (the screen, `SpanStyle`, cross-row selection, the growing
  composer).**
-  - [ ] **Android integration for this screen does not exist yet.** No
-    cdylib/Gradle shell the way `iris-android-app` wraps `tabs-ui` (I2),
-    so `transcript-bench.sh`'s render-number pass condition against the
-    Compose baseline cannot be run. Needs: real `client-core::ApiClient`/
-    `event_stream::follow_session_events` wiring against
-    `app/ui-sandbox.sh --delay` (this crate deliberately fetches nothing
-    itself, `transcript-ui/src/lib.rs`'s doc), a new cdylib + Gradle
-    module, then the bench script pointed at it.
+  - [x] **Android integration for this screen — done, 2026-09-05.**
+    `iris-android-app`'s `transcript-screen` Cargo feature
+    (`transcript_client.rs`) runs this screen against a real `ai-server`
+    through `client-core`, confirmed on-device: real scrolling, real
+    touch-drag panning, tap-by-name on the composer. Two real bugs found
+    and fixed along the way (a missing `INTERNET` permission; a
+    background-thread redraw request that crashed via a `Looper`
+    requirement, fixed by routing through `View::post_delayed` — see
+    `IRIS.md`'s `Tasks::redraw_handle` entry). See RUST.md's I5 box,
+    "The Android integration, done 2026-09-05" for the full account.
+  - [x] **A render-time number for iris, comparable to Compose's
+    `transcript-bench.sh` report — instrumentation done and a real number
+    obtained, 2026-09-05 (later the same day); the clean comparable loop
+    is not.** `iris_core::FrameReport` (`iris/core/src/render/
+    frame_report.rs`, `IRIS.md`'s new entry) times every frame from
+    `render()`'s redraw start to after `queue.submit`+`present()`, exposed
+    as two named on-screen controls ("Frame report", "Reset frame
+    report"). Driven against a real on-device touch-drag it read
+    `frames=34 janky%=61.76 p50=26.5ms p90=48.0ms p99=98.1ms
+    worst=98.1ms` — real, not inferred, but accumulated across several
+    gestures rather than one clean 24-swipe loop, because of the new
+    finding below. See RUST.md's I5 box, "Update, 2026-09-05, later the
+    same day" for the full account.
+  - [ ] **New, 2026-09-05: intermittent touch delivery to iris's
+    `SurfaceView` under this checkout's `EMU_GPU=software` emulator.**
+    The same swipe coordinates, confirmed (by scanning a screenshot
+    column for the first non-black pixel) to sit over real row text,
+    sometimes produced 30+ real frames and a screenshot diff and
+    sometimes produced zero of either, across otherwise-identical
+    `ui-trace` invocations. Not the already-understood "already at that
+    scroll edge" case (reproduced with content confirmed taller than the
+    viewport, in both directions). Leading candidate, not yet confirmed:
+    this checkout's emulator was independently observed at ~78% of one
+    CPU core, continuously, while idle on-screen — SwiftShader's software
+    rasterisation is CPU-bound by design, and a synthetic touch competing
+    with that load for delivery is plausible but unmeasured *during* a
+    failing gesture (the standing rule against diagnosing from
+    after-the-fact measurements applies here). Needs a sampler (load,
+    `dumpsys input`, a `-i 0` `ui-trace` capture) running while a failing
+    gesture is driven, and ideally a comparison under `-gpu host` (real
+    Vulkan) to see whether it is specific to software rendering. This is
+    what blocks the clean, comparable 24-swipe loop above.
+  - [x] **Long-press-then-drag-to-select — confirmed on-device, 2026-09-05
+    (later the same day).** `ui-trace` gained a `holddrag X1 Y1 X2 Y2
+    HOLD_MS MOVE_MS` action (`emulator-tools`, additive, extends the same
+    `MotionEvent`/`injectInputEvent` mechanism `swipe` already used):
+    press, hold past `LONG_PRESS`, move, release, as one continuous touch.
+    Driven against a real row (`holddrag 300 1850 300 2050 600 300`) it
+    produced `iris selection: begin at row ...` then a sequence of
+    `iris selection: extend to row ...` log lines
+    (`transcript-ui/src/selection.rs`, a new small `log` dependency since
+    selection has no accessibility label of its own yet — see the next
+    item), and a screenshot taken right after shows the expected
+    highlighted selection spanning multiple rows. `DragArbiter`'s own
+    unit tests already covered this sequence against a synthetic clock;
+    this is the first time it has been driven by a real device touch.
  - [x] **Touch-drag panning over a row's own rendered text — done,
    2026-09-05.** `row.rs` used to register `CursorSense::click_or_drag()`
    on each row's `TextEdit` for cross-row selection; `TextEdit::draw`'s
@@ -203,6 +286,20 @@ order and what "done" looks like. Tick and date them in place.
    --workspace` and `cargo ndk` (both `iris` and `transcript-ui`) all
    clean; `run-headless.sh` screenshot byte-identical to before the
    change (38578 bytes). See RUST.md's I5 box, "Gap closed, 2026-09-05".
+  - [x] **Intermittent touch-scroll dropout — root-caused and fixed,
+    2026-09-05.** Not the coalesced-`ACTION_MOVE` hypothesis the earlier
+    pass suspected (ruled out): a gesture's `ACTION_DOWN` can land on a
+    row's own padding/gap or its header, which no `CursorSense` covers,
+    so `DragArbiter` never gets `press_start` and sits in `Idle`
+    (answers `Undecided` forever) for that whole gesture. Fixed via a new
+    `DragArbiter::is_idle()` that `Selection::drag`
+    (`transcript-ui/src/selection.rs`) checks to recover a missed press
+    on the next `Pressing` frame. Four new unit tests. See RUST.md's I5
+    box, "Touch-scroll dropout root-caused, 2026-09-05", for the trace and
+    what a peer session sharing this checkout's emulator mid-pass
+    prevented from being re-verified end-to-end (the aggregate
+    `iris-scroll.sh` three-run confirmation and a re-taken FrameReport
+    row) — a future pass should finish that once the emulator is free.
  - [ ] **Row-level accessibility names.** The composer carries
    `.label("Message")`; transcript rows do not carry a `.label()` of
    their own yet, so `Widgets::named()` (I4) does not include them —
@@ -244,6 +341,42 @@ order and what "done" looks like. Tick and date them in place.
  everything, the same way input is**. Whatever the mechanism, a widget
  that does not animate must pay nothing and import nothing for it.

+## Build (for the port)
+
+Widgets `RUST.md`'s "The port, in order (decided 2026-09-05)" needs and
+iris does not have yet, one entry per gap, named against the P-step that
+first needs it. Move an entry up to "Fix" or tick it in place once built;
+do not duplicate it there.
+
+- [ ] **A history-paging cushion measured in on-screen viewports, not a
+  row count.** (**P1**.) `iris::widget::List` has no equivalent of the
+  Compose app's `HISTORY_SCREENS` — AGENTS.md's "Things that have
+  bitten" is explicit that a fixed row count under-fills a screen on a
+  tool-heavy transcript and over-fills one on a text-heavy one, so
+  whatever loads the next page has to ask the list how many viewports
+  are actually on screen, not assume a constant.
+- [ ] **A scaled thumbnail/image widget for an in-transcript image.**
+  (**P1**.) `SessionImage.kt`'s bitmap decode-and-downscale has no iris
+  counterpart; iris's own image widget (used by `bench_images.rs`) draws
+  a loaded texture but does nothing about sourcing or scaling one from a
+  server-produced attachment.
+- [ ] **A modal/dialog primitive.** (**P1**, reused by **P3** and
+  **P5**.) Needed for the session settings dialog, `UsageDialog`'s
+  equivalent, and the delete-with-`deleteForeign` confirmation with its
+  toggle switch. Build once, wherever it is first needed, rather than
+  once per screen that wants one.
+- [ ] **A horizontal gauge/bar widget.** (**P1**.) For
+  `SessionUsageBar`'s equivalent — a bounded fill reflecting a fraction,
+  nothing fancier.
+- [ ] **A `BusyItem` equivalent: a dimmed row carrying an operation
+  label that does not block its list's own scroll/drag.** (**P3**.) The
+  Compose version tried an overlay first and it swallowed the drag along
+  with the tap (AGENTS.md's "Shared appearance") — worth not repeating
+  that attempt in iris before building the row-level version directly.
+- [ ] **A toggle switch.** (**P3**.) For the delete dialog's
+  `deleteForeign` control; iris has no switch/checkbox widget yet as far
+  as this pass found.
+
 ## Reconsider

 - [ ] **`WidgetView`.** Iris is unsure of it: what she wants is an easy way
@@ -147,12 +147,37 @@ turn.

 Spawn: `claude -p --verbose --input-format stream-json --output-format
 stream-json --permission-mode <mode>` in the chosen working directory, plus
-`--model`. Wire-format notes are pinned against CLI 2.1.237 in
-`session/claude.rs`'s module doc: permissions need the hidden
-`--permission-prompt-tool stdio` flag, AskUserQuestion answers ride
-`updatedInput.answers` keyed by question text, and `set_model`/`interrupt`
+`--model` and, where one has been chosen, `--effort`. Wire-format notes are
+pinned against CLI 2.1.237 in `session/claude.rs`'s module doc: permissions
+need the hidden `--permission-prompt-tool stdio` flag, AskUserQuestion answers
+ride `updatedInput.answers` keyed by question text, and `set_model`/`interrupt`
 are control requests.

+**The thinking level is settled at launch** (added 2026-09-04, because it is
+the largest saving available on a long session: output is about an eighth of
+what a session costs and thinking is the bulk of output, against the ~1.5% that
+is prose). The CLI's only two setting control requests are `set_model` and
+`set_permission_mode` -- checked against the 2.1.258 binary -- so there is no
+way to ask a running process to think differently. `set_session_effort` is
+therefore shaped like `set_session_cwd` rather than like `set_session_model`:
+it records the level and **stops the process**, and the next message or Start
+launches one that has it. It lives in the session settings dialog beside the
+working directory for that reason, not on the session bar beside the model and
+the mode, which do take effect mid-turn. `None` is a level in its own right --
+the CLI's own default -- so the picker can return to it; a level this app named
+as the default instead would be this app choosing one.
+
+**What a new session starts at is `Config::default_effort`**, applied in
+`spawn_session` rather than filled in by the spawn screen, so it holds for an
+import and a bare API call as well. It is set by the spawn screen's own
+picker, whose label says so: one control, where new sessions are made, rather
+than a settings page for a single value. It is not on a provider, because
+providers are discovered and the next rediscovery would erase it, and not on
+the phone, because a second device would then spawn at a level nobody there
+chose. `GET`/`POST /defaults` carry it, as a struct rather than a bare value
+so the permission mode -- still hardcoded to `auto` on the spawn screen -- can
+move there without a second route.
+
 **`--resume` only ever runs when nothing else has that session open.** That
 is the rule behind the import refusal, the single `ClaudeDriver::launch`
 entry point, and the `Exited` correction below; two CLIs on one session file
@@ -169,10 +194,33 @@ deliberate and easy to undo by accident:
  when the process restarts. That leaves the Claude driver as the odd one
  out rather than this one — the CLI's memory is a cache in front of the same
  transcript. Resolve any inconsistency in this direction.
- **A llama session on an ssh host is refused.** The model is reached over
-  HTTP and forwarding that port is not built, so refusing beats silently
-  talking to the wrong machine. A transport is "run this" plus "reach this
-  port", and only the first half exists.
+- **A llama session runs on whatever machine its setup names** (2026-09-04,
+  the last of phase 5). A transport is "run this" plus "reach this port", and
+  the second half is `Transport::reserve_port` — the port the server binds
+  *there* and the port that reaches it *here*, the same number locally —
+  carried by `Launch::reaching` onto the connection that already runs the
+  command. `llama-server` binds loopback on the far machine, so nothing is
+  served to its network. The far port is a guess from a range below the
+  ephemeral one, because no portable way to ask a machine for a free port
+  avoids racing the bind anyway; a collision is not silent, since the server
+  fails to bind and the readiness poll reports what its log said.
+- **The model file lives on the machine that serves it** (2026-09-04). Each
+  setup names its own models directory (`SshConfig::models_dir`, default
+  `~/.local/share/ai-app/models` expanded *there*), and a spawn resolves the
+  key on that machine — one round trip answering "at /abs/path" or "missing",
+  so a model that is not there is refused at the spawn rather than becoming a
+  server that never becomes ready. The spawn screen offers
+  `GET /setups/{id}/models`, that machine's list, rather than `GET /models`,
+  which is this backend's downloads. Downloading *to* another machine is
+  deliberately not built: a multi-gigabyte transfer with no progress
+  anywhere, and the file gets there however anything else on that machine
+  did.
+- **The readiness poll watches the process, not only the port.** A model that
+  will not load, a port already taken, a flag an older build does not know:
+  all exit within a second and none will ever answer `/health`, so waiting
+  out the 300s timeout turned the server's own account of the problem into
+  "gave up". The failure carries the tail of `llama-server.log`, which on a
+  remote session is the only copy anybody reading the phone can see.

 ### Models (2026-08-28)

@@ -203,6 +251,13 @@ deliberate and easy to undo by accident:
  A driver says what to run; something above it turns that into a process.
  Otherwise transport knowledge sits inside a translator whose job is a wire
  format, and every future driver has to remember to do the same.
+- **A forwarded launch gets a pty and every other one does not** (measured
+  2026-09-04). Killing the ssh client ends a CLI because it closes the stdin
+  that CLI is reading; `llama-server` never reads its stdin, so the same kill
+  left it running on the far machine with the model loaded — one orphan per
+  stopped session. With `-tt` the far side takes SIGHUP when the connection
+  goes. Its log then arrives through a line discipline, which nothing parses.
+  `-T` stays everywhere else, where a pty would rewrite the JSONL.
 - **`command -v` follows ssh's non-login PATH**, which is narrower than an
  interactive shell's, so a binary somewhere unusual is invisible to
  discovery. Point `command` at an absolute path.
@@ -499,6 +554,25 @@ rate-limited bucket). Poll at ≥180 s, only while a Claude session exists or
 the usage screen is open, and cache the last answer. It is undocumented, so
 `usage.rs` treats every field as optional and degrades rather than erroring.

+**Per provider, not per machine (2026-09-04).** A machine is not what is
+metered; the provider a session runs is. One machine offers echo, the Claude
+CLI and a local model side by side, and only the second spends anything — so
+pairing a session with a snapshot by machine alone drew the CLI's five-hour
+window under every echo session on it, a quota that session cannot spend. A
+session now names its meter (`usageProvider`, from
+`DriverKind::usage_provider`, which `usage::providers_for` reads too, so the
+two lists cannot disagree) and `GET /usage` is matched on machine *and*
+provider. `None` is a session that meters nothing, and the phone draws
+nothing at all for it — not a zero, and not "unknown".
+
+`DriverKind::Echo` names a meter of its own that exists only when a test has
+asked for one: `/usage` in an echo session sets an invented answer
+(`usage::Fixture`), and with none set there is no snapshot and no bar. That
+is what makes those screens' states reachable — a number near the top, a
+window between blocks with no reset time, a machine nobody logged into, one
+that could not be reached — without spending real quota to arrange them,
+which is why none of them had ever been looked at.
+
 **Per machine, not per backend (2026-08-29).** The credential store that
 matters is the one on the machine the session runs on, because that is the
 account being billed — and in the layout this aims at, `ai-server` is on the
@@ -522,6 +596,68 @@ always running. So absent means **not running**, and only a timestamp that
 arrives and cannot be parsed is unknown. `WindowEnd` in `ResetCountdown.kt`
 is the one rule both readers go through.

+### Auto-resume (2026-09-05)
+
+**A session may pick itself back up when the account's usage limit lifts.**
+Off unless somebody switched that session to it, because it spends quota the
+moment quota exists and does so with nobody looking — that is not a thing a
+default may decide. It sends one message, `continue` unless another was
+typed, and then it is done; there is no retry loop around the conversation
+itself.
+
+**Running out of quota is a state, not an error.** `Event::LimitReached`
+carries the dialect's reset time where it gave one, and recognising it
+belongs to the driver — the Claude CLI ends the turn with `is_error` and
+`Claude AI usage limit reached|1788546972`, and nothing above the driver
+matches on a string. The transcript draws it as a divider, like a clear or a
+compaction: what a reader scrolling back wants from it is why the
+conversation stops at that line.
+
+**The schedule is a plan to ask, never a plan to send.** Every reset time
+available here is untrustworthy in the direction that matters: the dialect's
+is written when the turn fails, and the endpoint's moves when the window
+does. So the wait ends in a question to `usage.rs`, and only `ok` with no
+window at 100% sends anything. A window still spent reschedules to *its own*
+reset time — which is what makes a limit that lifts later than promised wait
+longer, and one that lifts sooner resume sooner. A meter that cannot be
+asked at all is a longer wait too, never a send: "we could not find out"
+must not be able to produce the same action as "there is room".
+
+Bounded, because something has to be: a day after the limit was hit the wait
+stops and says so in the session's own transcript. A machine that can never
+be asked would otherwise be retried for ever with nothing on screen saying
+so.
+
+The schedule is persisted on the session (`resume: Some(ScheduledResume)`),
+not held in memory: a five-hour window routinely outlasts a backend restart,
+and a wait forgotten across one is a session that silently never comes back.
+`resume.rs` is the top layer — it holds the manager and the monitor and
+neither holds it — which is what lets the decision be a pure function of a
+snapshot and a clock. The pump reports limits downward on a broadcast, for
+the reason `Shared` exists: the pump runs underneath the manager.
+
+**Exercised with echo, never with a real account.** `/limit [minutes]` in an
+echo session reports the same event a real driver does, and `/usage` sets
+what the meter answers — deliberately two commands, because the two
+disagreeing is the state the whole design is about. The loop was driven end
+to end that way on 2026-09-05: the wait moved from the dialect's two minutes
+to the meter's seven when the meter changed its mind, and the message went
+out on the first check after the meter came back under the limit.
+
+### Subagents (2026-09-05)
+
+**A subagent is a second transcript owned by a session, in the same event
+model, with no process and no controls of its own.** Full design and wire
+shape in `SUBAGENTS.md`, kept separate because the app half is being built
+against it in parallel and it is the shared contract between the two. The
+one-paragraph reason: a session's Task-tool helpers already speak the common
+event model on the parent's own stdout (each line carrying
+`parent_tool_use_id`), so giving each one its own small transcript — same
+file format, same paging routes, same SSE stream, reused by addressing rather
+than by copying — costs a routing step in the translator and a registry
+(`session/subagent.rs`) rather than a second session type with a driver, a
+process and a config entry it does not need.
+
 ### HTTP surface

 **`routes.rs`'s module doc comment is the table.** REST for actions, one SSE
@@ -800,8 +936,9 @@ Noticed and deliberately not fixed, so they are not re-found from scratch.

 Phases 1–3 (the skeleton pipe, the full Claude driver, the usage screen) done
 2026-08-24. Phase 4 (llama.cpp: model browsing, downloads, and `llama-server`
-through its OpenAI-compatible endpoint) and phase 5 (ssh) done 2026-08-28.
-The file explorer and the transcript cache followed in September. What is
+through its OpenAI-compatible endpoint) and phase 5 (ssh) done 2026-08-28,
+except for the remote `llama-server` and its port forward, which landed
+2026-09-04. The file explorer and the transcript cache followed in September. What is
 left is real-phone/WireGuard bring-up, which is operational rather than code.

 Each phase ended runnable and verified against the real thing. The backend
@@ -3651,6 +3651,7 @@ dependencies = [
 "client-core",
 "event-model",
 "iris",
+ "log",
 "pulldown-cmark",
 ]

@@ -57,6 +57,16 @@ send_wrapper = "0.6.0"
 # installs it -- this crate never installs a logger itself.
 log = "0.4.28"

+[features]
+# RUST.md's I5 "Where iris's frame time goes" diagnosis: forces the Android
+# `wgpu::Instance` to `Backends::GL` instead of `Backends::PRIMARY`, so the
+# same build can be measured against SwiftShader's software Vulkan ICD (the
+# default) or virgl's GLES path, without a second env-var plumbing path that
+# nothing on this machine can hand to an already-launched Android process
+# (there is no `am start` environment and no system-property reader here to
+# add one). Android-only; `android/render.rs` is the only reader.
+force-gles = []
+
 [dev-dependencies]
 tokio = { workspace = true, features = ["sync", "rt", "rt-multi-thread", "time"] }
 # The tabs example's widget tree. A dev-dependency cycle back to this
@@ -558,6 +558,12 @@ dependencies = [
 "arrayvec",
 ]

+[[package]]
+name = "base64"
+version = "0.23.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ac07cdecf99051d9a5238b80f35af32cdeba5b336e55d957b318b50137e18da5"
+
 [[package]]
 name = "bit-set"
 version = "0.8.0"
@@ -734,6 +740,16 @@ version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f079e83a288787bcd14a6aea84cee5c87a67c5a3e660c30f557a3d24761b3527"

+[[package]]
+name = "client-core"
+version = "0.1.0"
+dependencies = [
+ "event-model",
+ "serde",
+ "serde_json",
+ "ureq",
+]
+
 [[package]]
 name = "clipboard-win"
 version = "5.4.1"
@@ -779,6 +795,35 @@ dependencies = [
 "crossbeam-utils",
 ]

+[[package]]
+name = "cookie"
+version = "0.18.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1a373e3602691c3cdea496d2f0ee5935151e6168fe87739483c463db1b2f2f87"
+dependencies = [
+ "percent-encoding",
+ "time",
+ "version_check",
+]
+
+[[package]]
+name = "cookie_store"
+version = "0.22.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "15b2c103cf610ec6cae3da84a766285b42fd16aad564758459e6ecf128c75206"
+dependencies = [
+ "cookie",
+ "document-features",
+ "idna",
+ "indexmap",
+ "log",
+ "serde",
+ "serde_derive",
+ "serde_json",
+ "time",
+ "url",
+]
+
 [[package]]
 name = "core-foundation"
 version = "0.9.4"
@@ -886,6 +931,12 @@ version = "1.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f27ae1dd37df86211c42e150270f82743308803d90a6f6e6651cd730d5e1732f"

+[[package]]
+name = "deranged"
+version = "0.5.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7cd812cc2bc1d69d4764bd80df88b4317eaef9e773c75226407d9bc0876b211c"
+
 [[package]]
 name = "dispatch"
 version = "0.2.0"
@@ -1048,6 +1099,14 @@ dependencies = [
 "pin-project-lite",
 ]

+[[package]]
+name = "event-model"
+version = "0.1.0"
+dependencies = [
+ "serde",
+ "serde_json",
+]
+
 [[package]]
 name = "exr"
 version = "1.74.2"
@@ -1179,6 +1238,15 @@ version = "0.3.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "aa9a19cbb55df58761df49b23516a86d432839add4af60fc256da840f66ed35b"

+[[package]]
+name = "form_urlencoded"
+version = "1.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cb4cb245038516f5f85277875cdaa4f7d2c9a0fa0468de06ed190163b1581fcf"
+dependencies = [
+ "percent-encoding",
+]
+
 [[package]]
 name = "futures-core"
 version = "0.3.34"
@@ -1253,6 +1321,26 @@ dependencies = [
 "windows-link",
 ]

+[[package]]
+name = "getopts"
+version = "0.2.24"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cfe4fbac503b8d1f88e6676011885f34b7174f46e59956bba534ba83abded4df"
+dependencies = [
+ "unicode-width",
+]
+
+[[package]]
+name = "getrandom"
+version = "0.2.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ff2abc00be7fca6ebc474524697ae276ad847ad0a6b3faa4bcb027e9a4614ad0"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "wasi",
+]
+
 [[package]]
 name = "getrandom"
 version = "0.3.4"
@@ -1423,6 +1511,22 @@ version = "0.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "dfa686283ad6dd069f105e5ab091b04c62850d3e4cf5d67debad1933f55023df"

+[[package]]
+name = "http"
+version = "1.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "918d3568bebf352712bc2ef3d46a8bcf1a75b373be6539de198e9105cbbf9ce0"
+dependencies = [
+ "bytes",
+ "itoa",
+]
+
+[[package]]
+name = "httparse"
+version = "1.10.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6dbf3de79e51f3d586ab4cb9d5c3e2c14aa28ed23d180cf89b4df0454a69cc87"
+
 [[package]]
 name = "icu_collections"
 version = "2.3.0"
@@ -1551,6 +1655,27 @@ version = "2.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ae293c039020f9ec10710af98d29ce6aa2051486638b49c9a6409f3b4a9e98ad"

+[[package]]
+name = "idna"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3b0875f23caa03898994f6ddc501886a45c7d3d62d04d2d90788d47be1b1e4de"
+dependencies = [
+ "idna_adapter",
+ "smallvec",
+ "utf8_iter",
+]
+
+[[package]]
+name = "idna_adapter"
+version = "1.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cb68373c0d6620ef8105e855e7745e18b0d00d3bdb07fb532e434244cdb9a714"
+dependencies = [
+ "icu_normalizer",
+ "icu_properties",
+]
+
 [[package]]
 name = "image"
 version = "0.25.10"
@@ -1640,9 +1765,15 @@ version = "0.1.0"
 dependencies = [
 "android-view",
 "android_logger",
+ "client-core",
+ "event-model",
 "iris",
+ "libc",
 "log",
+ "serde_json",
 "tabs-ui",
+ "tokio",
+ "transcript-ui",
 ]

 [[package]]
@@ -1676,6 +1807,12 @@ dependencies = [
 "either",
 ]

+[[package]]
+name = "itoa"
+version = "1.0.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
+
 [[package]]
 name = "jni"
 version = "0.21.1"
@@ -2096,6 +2233,12 @@ dependencies = [
 "num-traits",
 ]

+[[package]]
+name = "num-conv"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "521739c6d2bac4aa25192232afe6841231376b2b26d4d9fae5ecf8ca5772e441"
+
 [[package]]
 name = "num-derive"
 version = "0.4.2"
@@ -2744,6 +2887,12 @@ dependencies = [
 "zerovec",
 ]

+[[package]]
+name = "powerfmt"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "439ee305def115ba05938db6eb1644ff94165c5ab5e9420d1c1bcedbba909391"
+
 [[package]]
 name = "ppv-lite86"
 version = "0.2.21"
@@ -2796,6 +2945,25 @@ dependencies = [
 "syn 2.0.119",
 ]

+[[package]]
+name = "pulldown-cmark"
+version = "0.13.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e9f068eba8e7071c5f9511831b44f32c740d5adf574e990f946ddb53db2f314e"
+dependencies = [
+ "bitflags 2.13.1",
+ "getopts",
+ "memchr",
+ "pulldown-cmark-escape",
+ "unicase",
+]
+
+[[package]]
+name = "pulldown-cmark-escape"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "007d8adb5ddab6f8e3f491ac63566a7d5002cc7ed73901f72057943fa71ae1ae"
+
 [[package]]
 name = "pulp"
 version = "0.22.3"
@@ -3075,6 +3243,20 @@ version = "0.8.53"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "47b34b781b31e5d73e9fbc8689c70551fd1ade9a19e3e28cfec8580a79290cc4"

+[[package]]
+name = "ring"
+version = "0.17.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a4689e6c2294d81e88dc6261c768b63bc4fcdb852be6d1352498b114f61383b7"
+dependencies = [
+ "cc",
+ "cfg-if",
+ "getrandom 0.2.17",
+ "libc",
+ "untrusted",
+ "windows-sys 0.52.0",
+]
+
 [[package]]
 name = "roxmltree"
 version = "0.21.1"
@@ -3125,6 +3307,41 @@ dependencies = [
 "windows-sys 0.61.2",
 ]

+[[package]]
+name = "rustls"
+version = "0.23.43"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0283386ce02abc0151e1761d08802dfe86c173b0b494af5cbc086574e453da06"
+dependencies = [
+ "log",
+ "once_cell",
+ "ring",
+ "rustls-pki-types",
+ "rustls-webpki",
+ "subtle",
+ "zeroize",
+]
+
+[[package]]
+name = "rustls-pki-types"
+version = "1.15.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2f4925028c7eb5d1fcdaf196971378ed9d2c1c4efc7dc5d011256f76c99c0a96"
+dependencies = [
+ "zeroize",
+]
+
+[[package]]
+name = "rustls-webpki"
+version = "0.103.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f3c3cf1d8b1e7d4927e2d154c3fcb02979afb9939629c62cd9048d4f07b60ac2"
+dependencies = [
+ "ring",
+ "rustls-pki-types",
+ "untrusted",
+]
+
 [[package]]
 name = "rustversion"
 version = "1.0.23"
@@ -3207,6 +3424,19 @@ dependencies = [
 "syn 3.0.5",
 ]

+[[package]]
+name = "serde_json"
+version = "1.0.151"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c841b55ecdae098c80dcae9cf767f6f8a0c2cdb3416bbef72181df4d0fe73f14"
+dependencies = [
+ "itoa",
+ "memchr",
+ "serde",
+ "serde_core",
+ "zmij",
+]
+
 [[package]]
 name = "serde_repr"
 version = "0.1.21"
@@ -3363,6 +3593,12 @@ version = "0.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6637bab7722d379c8b41ba849228d680cc12d0a45ba1fa2b48f2a30577a06731"

+[[package]]
+name = "subtle"
+version = "2.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "13c2bddecc57b384dee18652358fb23172facb8a2c51ccc10d74c157bdea3292"
+
 [[package]]
 name = "swash"
 version = "0.2.10"
@@ -3490,6 +3726,36 @@ dependencies = [
 "zune-jpeg",
 ]

+[[package]]
+name = "time"
+version = "0.3.55"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cdb87b95ec50ddfa440816d227a17b2ccbdda963a316a727fda0fc4334f7d134"
+dependencies = [
+ "deranged",
+ "num-conv",
+ "powerfmt",
+ "serde_core",
+ "time-core",
+ "time-macros",
+]
+
+[[package]]
+name = "time-core"
+version = "0.1.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9e1c906769ad99c88eaa54e728060edef082f8e358ff32030cb7c7d315e81109"
+
+[[package]]
+name = "time-macros"
+version = "0.2.32"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7e689342a48d2ea927c87ea50cabf8594854bf940e9310208848d680d668ed85"
+dependencies = [
+ "num-conv",
+ "time-core",
+]
+
 [[package]]
 name = "tiny-skia"
 version = "0.11.4"
@@ -3596,6 +3862,17 @@ dependencies = [
 "once_cell",
 ]

+[[package]]
+name = "transcript-ui"
+version = "0.1.0"
+dependencies = [
+ "client-core",
+ "event-model",
+ "iris",
+ "log",
+ "pulldown-cmark",
+]
+
 [[package]]
 name = "tree_magic_mini"
 version = "3.2.2"
@@ -3634,6 +3911,12 @@ dependencies = [
 "keyboard-types",
 ]

+[[package]]
+name = "unicase"
+version = "2.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dbc4bc3a9f746d862c45cb89d705aa10f187bb96c76001afab07a0d35ce60142"
+
 [[package]]
 name = "unicode-ident"
 version = "1.0.24"
@@ -3652,6 +3935,62 @@ version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b4ac048d71ede7ee76d585517add45da530660ef4390e49b098733c6e897f254"

+[[package]]
+name = "untrusted"
+version = "0.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8ecb6da28b8a351d773b68d5825ac39017e680750f980f3a1a85cd8dd28a47c1"
+
+[[package]]
+name = "ureq"
+version = "3.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "972d7902c8735f2695410b8aed7df6ed12a47394aa1c8d7af49f0497b731a94d"
+dependencies = [
+ "base64",
+ "cookie_store",
+ "flate2",
+ "log",
+ "percent-encoding",
+ "rustls",
+ "rustls-pki-types",
+ "serde",
+ "serde_json",
+ "ureq-proto",
+ "utf8-zero",
+ "webpki-roots",
+]
+
+[[package]]
+name = "ureq-proto"
+version = "0.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "da5f78b09e6941e1a0f2e30e695e4b120377b54d5e0aec11b594bb57b3971613"
+dependencies = [
+ "base64",
+ "http",
+ "httparse",
+ "log",
+]
+
+[[package]]
+name = "url"
+version = "2.5.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ff67a8a4397373c3ef660812acab3268222035010ab8680ec4215f38ba3d0eed"
+dependencies = [
+ "form_urlencoded",
+ "idna",
+ "percent-encoding",
+ "serde",
+]
+
+[[package]]
+name = "utf8-zero"
+version = "0.8.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b8c0a043c9540bae7c578c88f91dda8bd82e59ae27c21baca69c8b191aaf5a6e"
+
 [[package]]
 name = "utf8_iter"
 version = "1.0.4"
@@ -3696,6 +4035,12 @@ dependencies = [
 "winapi-util",
 ]

+[[package]]
+name = "wasi"
+version = "0.11.1+wasi-snapshot-preview1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ccf3ec651a847eb01de73ccad15eb7d99f80485de043efb2f370cd654f4ea44b"
+
 [[package]]
 name = "wasip2"
 version = "1.0.4+wasi-0.2.12"
@@ -3889,6 +4234,15 @@ dependencies = [
 "wasm-bindgen",
 ]

+[[package]]
+name = "webpki-roots"
+version = "1.0.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7dcd9d09a39985f5344844e66b0c530a33843579125f23e21e9f0f220850f22a"
+dependencies = [
+ "rustls-pki-types",
+]
+
 [[package]]
 name = "weezl"
 version = "0.1.12"
@@ -4748,6 +5102,12 @@ dependencies = [
 "synstructure",
 ]

+[[package]]
+name = "zeroize"
+version = "1.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e13c156562582aa81c60cb29407084cdb54c4164760106ab78e6c5b0858cf64e"
+
 [[package]]
 name = "zerotrie"
 version = "0.2.5"
@@ -4789,6 +5149,12 @@ version = "0.6.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "34b31d188d9d685a4f9c7b46d6e36631b07058d2cfe190267adce54dc230bf12"

+[[package]]
+name = "zmij"
+version = "1.0.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "29666d0abbfad1e3dc4dcf6144730dd3a3ab225bbbdac83319345b1b44ccfc1b"
+
 [[package]]
 name = "zune-core"
 version = "0.5.3"
@@ -17,10 +17,55 @@ crate-type = ["cdylib"]

 [dependencies]
 iris = { path = "../" }
-tabs-ui = { path = "../tabs-ui" }
 android-view = { git = "https://github.com/rust-mobile/android-view.git", rev = "bec6c62a96cef8239b0fd7fedeef9b184d02e3a1" }
 android_logger = "0.15.0"
 log = "0.4.28"
+# `tabs-screen` (default, I2/I4's demo) and `transcript-screen` (I5's
+# Android integration) are mutually exclusive -- one `ActiveClient` type is
+# compiled in, never both (`lib.rs`'s doc comment) -- so both sets of deps
+# are optional and each screen's feature pulls in only its own. Without
+# this, building `--features transcript-screen` alone (default features
+# still on) left `tabs-ui` linked but never referenced under that cfg,
+# which Cargo's `unused_dependencies` lint (on by default) correctly flags.
+tabs-ui = { path = "../tabs-ui", optional = true }
+transcript-ui = { path = "../transcript-ui", optional = true }
+client-core = { path = "../../client-core", optional = true }
+event-model = { path = "../../event-model", optional = true }
+serde_json = { version = "1", features = ["float_roundtrip"], optional = true }
+# P0's bench build only (docs/RUST.md): `getrusage(RUSAGE_SELF)` for
+# process CPU time, matching `libc::getrusage`'s mention in that box over
+# parsing `/proc/self/stat` by hand and assuming `USER_HZ`. Already in the
+# workspace's own dependency tree transitively (`iris/Cargo.lock`, pinned
+# at 0.2.179) -- this makes it a direct dependency at the same version
+# rather than a second, possibly-drifting resolution.
+libc = { version = "0.2.179", optional = true }
+# P0's bench build only: the scroll animation and the streaming phase are
+# both a sequence of `sleep`s inside the async task `rsc.spawn_task` already
+# runs on iris's own tokio runtime (`iris/src/task.rs`'s `Tasks::init`), and
+# the battery sampler is a second, concurrent task on that same runtime
+# (`tokio::spawn`) -- so this crate needs `tokio` directly rather than only
+# through `iris`. `rt`+`time` only: no I/O, no macros, nothing this crate
+# doesn't call. Version matches the one `iris`'s own dependency tree already
+# resolves to (`iris/Cargo.lock`), so there is one copy of the runtime, not
+# two.
+tokio = { version = "1.53.1", features = ["rt", "time"], optional = true }
+
+[features]
+default = ["tabs-screen"]
+tabs-screen = ["dep:tabs-ui"]
+transcript-screen = ["dep:transcript-ui", "dep:client-core", "dep:event-model", "dep:serde_json"]
+# RUST.md's I5 "Where iris's frame time goes": forces the GLES backend
+# instead of SwiftShader's software Vulkan. See `iris/Cargo.toml`'s own doc
+# on the feature this forwards to.
+force-gles = ["iris/force-gles"]
+# P0's iris half (docs/RUST.md, docs/AGENTS.md's "The rigs"): the same
+# checked-in fixture, scroll loop and streaming phase the Compose `bench`
+# build type drives, run here against `transcript-ui`'s real screen with no
+# server. Depends on `transcript-screen` for `transcript-ui`/`client-core`/
+# `event-model` -- `lib.rs`'s `ActiveClient` selection gives this feature
+# priority over `transcript-screen`'s own `TranscriptClient` when both are
+# listed, which is how this crate's build command names both explicitly.
+bench = ["transcript-screen", "dep:libc", "dep:tokio"]

 [profile.release]
 panic = "abort"
@@ -19,9 +19,39 @@ android {
        versionName = "1.0"
    }

+    // A release build must be signed, and the key is per machine rather than per repo -- same
+    // reasoning and the same key as `app/build-apk.sh` (the Compose app): it is what a phone
+    // recognises the app by, and a secret never lives in a checkout (the mount is shared with an
+    // untrusted VM). `build-apk.sh` generates this key once and points at it through the
+    // environment; without it a release build here is unsigned, which is fine for everything
+    // except installing.
+    def keystore = System.getenv("AI_APP_KEYSTORE")
+    signingConfigs {
+        if (keystore != null) {
+            release {
+                storeFile = file(keystore)
+                storePassword = System.getenv("AI_APP_KEYSTORE_PASSWORD")
+                keyAlias = "ai-app"
+                keyPassword = storePassword
+            }
+        }
+    }
+
    buildTypes {
        debug {
        }
+        // P0's iris half (docs/RUST.md's P0 box): the build a phone actually runs. The `.so`
+        // itself is built separately with `cargo ndk --release --features "transcript-screen
+        // force-gles bench"` straight into src/main/jniLibs/ (this crate's own Cargo.toml) --
+        // Gradle here only packages and signs whatever is already there, the same division as the
+        // debug/tabs-screen build this project started with. `applicationIdSuffix` keeps it
+        // installable beside a debug build of the tabs demo rather than replacing it.
+        release {
+            applicationIdSuffix ".bench"
+            if (keystore != null) {
+                signingConfig = signingConfigs.release
+            }
+        }
    }

    compileOptions {
@@ -1,6 +1,15 @@
 <?xml version="1.0" encoding="utf-8"?>
 <manifest xmlns:android="http://schemas.android.com/apk/res/android">

+    <!-- Only needed by the transcript-screen feature (RUST.md's I5),
+         which talks to a real ai-server; the plain tabs demo (I2/I4) makes
+         no network call and never noticed this was missing. Absent,
+         UreqTransport::new's connect failed with EPERM (Operation not
+         permitted), not the ECONNREFUSED/ENETUNREACH a firewall or a dead
+         server would give: a seccomp-level socket denial reads nothing
+         like a network problem, which is what made it worth a comment. -->
+    <uses-permission android:name="android.permission.INTERNET" />
+
    <application
        android:allowBackup="true"
        android:label="iris android-view demo"
@@ -0,0 +1,100 @@
+// Only does anything under the `transcript-screen` feature (RUST.md's I5
+// Android integration) -- the plain tabs build (I2/I4) needs none of this
+// and stays untouched, same reasoning as the feature gate in Cargo.toml.
+//
+// Bakes the sandbox server's host, port, token and pinned CA in at build
+// time, the same way `app/androidApp/build.gradle.kts`'s
+// `GeneratePinnedCert` task bakes the CA for the Compose app -- see that
+// file's comment for why reading the machine's own certificate at build
+// time is the right trust boundary. This build additionally bakes the
+// host/port/token, which the Compose app does not: that app enrolls at
+// runtime from a scanned QR/deep link, and a from-scratch enrollment UI
+// (Keystore-sealed token storage, a QR/link scanner) is real, separate
+// scope this integration does not need to build to answer RUST.md's
+// question -- there is nothing here yet resembling `ServerConfig.kt`. So
+// this is a **deliberate simplification for this rig only**: an APK built
+// this way is good for exactly the emulator/server pair that built it, and
+// must never be treated as a template for a real enrollment flow. Recorded
+// in RUST.md's I5 box rather than left to be rediscovered.
+use std::path::PathBuf;
+
+fn main() {
+    if std::env::var_os("CARGO_FEATURE_TRANSCRIPT_SCREEN").is_none() {
+        return;
+    }
+    // P0's bench build (docs/RUST.md) opens the checked-in fixture with no
+    // server at all -- `bench_client.rs` never references the `pinned`
+    // module this generates, so requiring a live server's host/port/token/
+    // CA to build it (as plain `transcript-screen` does, below) would be a
+    // pointless requirement for a build that talks to nothing.
+    if std::env::var_os("CARGO_FEATURE_BENCH").is_some() {
+        return;
+    }
+    println!("cargo:rerun-if-env-changed=AI_APP_TRANSCRIPT_HOST");
+    println!("cargo:rerun-if-env-changed=AI_APP_TRANSCRIPT_PORT");
+    println!("cargo:rerun-if-env-changed=AI_APP_TRANSCRIPT_TOKEN");
+    println!("cargo:rerun-if-env-changed=AI_APP_CA");
+    println!("cargo:rerun-if-env-changed=XDG_CONFIG_HOME");
+
+    let host = require_env(
+        "AI_APP_TRANSCRIPT_HOST",
+        "the sandbox server's host as the emulator reaches it, e.g. 10.0.2.2",
+    );
+    let port = require_env(
+        "AI_APP_TRANSCRIPT_PORT",
+        "the sandbox server's port -- app/ui-sandbox.sh's start banner prints it",
+    );
+    let token = require_env(
+        "AI_APP_TRANSCRIPT_TOKEN",
+        "the bearer token -- ~/.config/ai-app/sandbox-token, or the start banner's enrollment link",
+    );
+
+    let ca_path = std::env::var_os("AI_APP_CA")
+        .map(PathBuf::from)
+        .unwrap_or_else(|| {
+            let base = std::env::var_os("XDG_CONFIG_HOME")
+                .map(PathBuf::from)
+                .unwrap_or_else(|| {
+                    let home = std::env::var_os("HOME").expect("HOME must be set");
+                    PathBuf::from(home).join(".config")
+                });
+            base.join("ai-app").join("certs").join("ca.pem")
+        });
+    let ca_pem = std::fs::read_to_string(&ca_path).unwrap_or_else(|e| {
+        panic!(
+            "no CA certificate at {} ({e}).\n\
+             Start ai-server (or app/ui-sandbox.sh) once on this machine first -- it \
+             generates the CA this build pins. Set AI_APP_CA=/path/to/ca.pem to build \
+             against a different one.",
+            ca_path.display()
+        )
+    });
+    let ca_pem = ca_pem.trim();
+    if !ca_pem.starts_with("-----BEGIN CERTIFICATE-----") {
+        panic!("{} is not a PEM certificate.", ca_path.display());
+    }
+
+    let out_dir = PathBuf::from(std::env::var_os("OUT_DIR").unwrap());
+    let generated = format!(
+        "// Generated by build.rs from {host}:{port} and {ca}. Do not edit.\n\
+         pub const HOST: &str = {host_lit:?};\n\
+         pub const PORT: u16 = {port};\n\
+         pub const TOKEN: &str = {token_lit:?};\n\
+         pub const CA_PEM: &str = {ca_lit:?};\n",
+        host = host,
+        port = port
+            .parse::<u16>()
+            .unwrap_or_else(|e| panic!("AI_APP_TRANSCRIPT_PORT={port:?} is not a u16: {e}")),
+        ca = ca_path.display(),
+        host_lit = host,
+        token_lit = token,
+        ca_lit = ca_pem,
+    );
+    std::fs::write(out_dir.join("pinned_config.rs"), generated).unwrap();
+}
+
+fn require_env(name: &str, what: &str) -> String {
+    std::env::var(name).unwrap_or_else(|_| {
+        panic!("{name} must be set to build the transcript-screen feature -- {what}")
+    })
+}
@@ -0,0 +1,418 @@
+//! P0's iris half (docs/RUST.md's P0 box, docs/AGENTS.md's "The rigs"):
+//! the same fixture, scroll loop and streaming phase the Compose `bench`
+//! build type's `BenchRun.kt`/`BenchFixture.kt` drive, run here against
+//! `transcript-ui`'s real screen with no server -- a frame-time comparison
+//! that measures the renderer rather than the data or the network.
+//!
+//! **Reuses `transcript_client.rs`'s shape** (folded items, a full
+//! `transcript_ui::build_tree` rebuild per event) with the network half
+//! replaced by the checked-in fixture, embedded with `include_str!` --
+//! `app/bench-fixture/assets/transcript.jsonl`, 1,915,760 bytes, generated
+//! by `app/bench-fixture/generate.py` and never a real transcript (that
+//! file's own README). The first 3,200 lines are the opening backlog,
+//! folded once through `client_core::transcript_fold::fold_page` exactly
+//! as a real `/transcript` page would be; the remaining ~400 are the
+//! streaming tail, replayed one at a time through `fold_event` -- the same
+//! fold path a live SSE reply arrives on -- by the "Run benchmark"
+//! control below.
+
+use crate::bench_jni::PlatformHandle;
+use android_view::jni::{JavaVM, objects::GlobalRef};
+use client_core::transcript_fold::{TranscriptItem, fold_event, fold_page, group_tool_runs};
+use event_model::SeqEvent;
+use iris::android::{AndroidAppState, AndroidRsc, AndroidUiState, HasAndroidUiState};
+use iris::prelude::*;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::time::Duration;
+
+/// bench-fixture/README.md: the first `BACKLOG_COUNT` non-blank lines are
+/// the opening window; the rest are the streaming tail. Kept in sync with
+/// `BenchFixture.kt`'s identical constant by hand -- both read the same
+/// checked-in file, so a mismatch would only mean the two apps' bench
+/// builds open a different split of it, not a wrong-vs-right answer.
+const BACKLOG_COUNT: usize = 3200;
+
+/// `BenchRun.kt`'s own constants -- kept identical so the two apps' bench
+/// runs are the same gesture and the same load, which is the entire point
+/// of a shared fixture and a shared scripted loop (P0's pass condition).
+const CYCLES: usize = 6;
+const SWIPE_PX: f32 = 900.0;
+const SWIPE_MS: u64 = 200;
+const SWIPE_PAUSE_MS: u64 = 500;
+const STREAM_EVENTS_PER_SEC: u64 = 20;
+const STREAM_SECONDS: u64 = 20;
+/// One animation step's target cadence -- close enough to 60Hz that a
+/// `List::scroll` swipe is many small moves rather than one jump, so
+/// frames are actually rendered along the way (the point of animating it
+/// at all rather than calling `scroll` once per swipe).
+const ANIM_STEP_MS: u64 = 16;
+
+const FIXTURE_JSONL: &str = include_str!("../../../app/bench-fixture/assets/transcript.jsonl");
+
+pub struct BenchClient {
+    ui_state: AndroidUiState,
+    content: WeakWidget<WidgetPtr>,
+    report_display: WeakWidget<TextEdit>,
+    screen: Option<transcript_ui::TranscriptScreen>,
+    items: Vec<TranscriptItem>,
+    /// The events not yet streamed -- consumed by `start_benchmark`'s own
+    /// clone, kept here only as the source a second run would need (the
+    /// button can be pressed more than once; `running` just stops overlap,
+    /// not repeat).
+    stream_tail: Vec<SeqEvent>,
+    platform: Option<Arc<PlatformHandle>>,
+    last_report: Option<String>,
+    running: bool,
+}
+
+impl HasAndroidUiState for BenchClient {
+    fn android_state(&self) -> &AndroidUiState {
+        &self.ui_state
+    }
+    fn android_state_mut(&mut self) -> &mut AndroidUiState {
+        &mut self.ui_state
+    }
+}
+
+/// Parses the fixture once: `serde_json::Value`s for the backlog
+/// (`fold_page` takes a page of raw wire JSON, same as a real
+/// `/transcript` response) and folded `SeqEvent`s for the tail (`fold_event`
+/// takes one live wire event at a time, same as a real SSE frame).
+fn parse_fixture() -> (Vec<serde_json::Value>, Vec<SeqEvent>) {
+    let lines: Vec<&str> = FIXTURE_JSONL
+        .lines()
+        .filter(|line| !line.trim().is_empty())
+        .collect();
+    let mut backlog = Vec::with_capacity(BACKLOG_COUNT.min(lines.len()));
+    let mut stream_tail = Vec::new();
+    for (i, line) in lines.iter().enumerate() {
+        let value: serde_json::Value =
+            serde_json::from_str(line).expect("bench fixture is generated JSON, always valid");
+        if i < BACKLOG_COUNT {
+            backlog.push(value);
+        } else {
+            let event: SeqEvent = serde_json::from_value(value)
+                .expect("bench fixture event matches event-model's SeqEvent");
+            stream_tail.push(event);
+        }
+    }
+    (backlog, stream_tail)
+}
+
+fn placeholder<Rsc: HasEvents>(rsc: &mut Rsc, message: &str) -> StrongWidget {
+    wtext(message.to_string())
+        .color(Color::WHITE)
+        .wrap(true)
+        .pad(16)
+        .add_strong(rsc)
+        .any()
+}
+
+/// `getrusage(RUSAGE_SELF)`'s user+system time, in ms -- `None` only if
+/// the syscall itself fails, which UI_RULES.md's "never present an
+/// inferred value as a measured one" says to keep apart from a real (and
+/// here, impossible) zero.
+fn process_cpu_ms() -> Option<u64> {
+    // SAFETY: `rusage` is a plain-old-data struct `getrusage` fully
+    // initialises on success; on failure it is never read.
+    unsafe {
+        let mut usage: libc::rusage = std::mem::zeroed();
+        if libc::getrusage(libc::RUSAGE_SELF, &mut usage) != 0 {
+            return None;
+        }
+        let user_ms = usage.ru_utime.tv_sec as u64 * 1000 + usage.ru_utime.tv_usec as u64 / 1000;
+        let sys_ms = usage.ru_stime.tv_sec as u64 * 1000 + usage.ru_stime.tv_usec as u64 / 1000;
+        Some(user_ms + sys_ms)
+    }
+}
+
+/// `VmHWM` from `/proc/self/status` -- the process's peak RSS since it
+/// started, in kB. Same source `BenchRun.kt`'s `peakRssLine` reads, so the
+/// two reports' numbers mean the same thing.
+fn peak_rss_kb() -> Option<u64> {
+    std::fs::read_to_string("/proc/self/status")
+        .ok()?
+        .lines()
+        .find_map(|line| line.strip_prefix("VmHWM:"))
+        .and_then(|rest| rest.trim().strip_suffix("kB"))
+        .and_then(|n| n.trim().parse().ok())
+}
+
+fn battery_line(samples: &[i32]) -> String {
+    if samples.is_empty() {
+        return "  battery current: unavailable on this device".to_string();
+    }
+    let mean = samples.iter().map(|&v| v as i64).sum::<i64>() / samples.len() as i64;
+    let min = samples.iter().min().unwrap();
+    let max = samples.iter().max().unwrap();
+    format!(
+        "  battery current: mean {mean}\u{b5}A over {} samples (min {min}, max {max})",
+        samples.len()
+    )
+}
+
+impl AndroidAppState for BenchClient {
+    fn new(mut ui_state: AndroidUiState, rsc: &mut AndroidRsc<Self>) -> Self {
+        let content = WidgetPtr::new().add(rsc);
+        let loading = placeholder(rsc, "Loading fixture...");
+        content(rsc).set(loading);
+
+        let report_display = wtext("")
+            .editable(EditMode::MultiLine)
+            .text_align(Align::LEFT)
+            .wrap(true)
+            .size(14)
+            .color(Color::WHITE)
+            .attr::<Selectable>(())
+            .label("Benchmark report")
+            .add(rsc);
+
+        let controls = bench_controls(rsc);
+        let tree = (
+            controls,
+            content.height(rest(2)),
+            report_display.height(rest(1)).pad(8),
+        )
+            .span(Dir::DOWN)
+            .add_strong(rsc)
+            .any();
+        ui_state.set_root(tree);
+
+        let mut client = Self {
+            ui_state,
+            content,
+            report_display,
+            screen: None,
+            items: Vec::new(),
+            stream_tail: Vec::new(),
+            platform: None,
+            last_report: None,
+            running: false,
+        };
+
+        let (backlog, stream_tail) = parse_fixture();
+        client.stream_tail = stream_tail;
+        match fold_page(&backlog) {
+            Ok(items) => {
+                client.items = items;
+                client.rebuild_transcript(rsc);
+            }
+            Err(message) => {
+                client.show_message(rsc, &format!("Couldn't fold the bench fixture: {message}"))
+            }
+        }
+        client
+    }
+
+    fn platform_ready(&mut self, _rsc: &mut AndroidRsc<Self>, vm: JavaVM, view: GlobalRef) {
+        self.platform = Some(Arc::new(PlatformHandle::new(vm, view)));
+    }
+
+    fn back_pressed(&mut self, _rsc: &mut AndroidRsc<Self>, _render: &mut UiRenderState) -> bool {
+        false
+    }
+}
+
+type Rsc = AndroidRsc<BenchClient>;
+
+fn bench_controls(rsc: &mut Rsc) -> WeakWidget {
+    let run_rect = rect(Color::rgb(40, 70, 40))
+        .on(
+            CursorSense::click(),
+            |ctx: EventIdCtx<'_, Rsc, _, _>, rsc: &mut Rsc| {
+                ctx.state.start_benchmark(rsc);
+            },
+        )
+        .label("Run benchmark");
+    let run = (
+        run_rect,
+        wtext("Run benchmark").size(18).text_align(Align::CENTER),
+    )
+        .stack()
+        .pad(8)
+        .add(rsc);
+
+    let copy_rect = rect(Color::rgb(50, 50, 60))
+        .on(
+            CursorSense::click(),
+            |ctx: EventIdCtx<'_, Rsc, _, _>, _rsc: &mut Rsc| {
+                ctx.state.copy_report();
+            },
+        )
+        .label("Copy report");
+    let copy = (
+        copy_rect,
+        wtext("Copy report").size(18).text_align(Align::CENTER),
+    )
+        .stack()
+        .pad(8)
+        .add(rsc);
+
+    (run, copy).span(Dir::RIGHT).height(56).add(rsc)
+}
+
+impl BenchClient {
+    fn show_message(&mut self, rsc: &mut Rsc, message: &str) {
+        let widget = placeholder(rsc, message);
+        (self.content)(rsc).set(widget);
+        self.screen = None;
+    }
+
+    fn rebuild_transcript(&mut self, rsc: &mut Rsc) {
+        let rows = group_tool_runs(&self.items);
+        let (screen, tree) = transcript_ui::build_tree(rsc, rows);
+        (self.content)(rsc).set(tree);
+        self.screen = Some(screen);
+    }
+
+    fn copy_report(&mut self) {
+        let Some(report) = &self.last_report else {
+            log::info!("iris bench report: nothing to copy -- run the benchmark first");
+            return;
+        };
+        let Some(platform) = &self.platform else {
+            log::info!("iris bench report: no platform handle, can't reach the clipboard");
+            return;
+        };
+        if platform.copy_to_clipboard("iris bench report", report) {
+            log::info!("iris bench report: copied to clipboard");
+        } else {
+            log::info!("iris bench report: clipboard copy failed");
+        }
+    }
+
+    /// P0's scripted run: `BenchRun.kt`'s scroll loop, then its streaming
+    /// phase, then the report -- run in-process for the same reason that
+    /// file's own doc gives (no usable system tracing on a real phone, no
+    /// agent that can drive one).
+    fn start_benchmark(&mut self, rsc: &mut Rsc) {
+        if self.running {
+            log::info!("iris bench report: already running");
+            return;
+        }
+        self.running = true;
+        self.android_state_mut().frame_report.reset();
+        self.report_display.edit(rsc).set("Running benchmark...");
+
+        let redraw = rsc.tasks.redraw_handle();
+        let platform = self.platform.clone();
+        let stream_tail = self.stream_tail.clone();
+        let cpu_start = process_cpu_ms();
+
+        rsc.spawn_task(async move |mut ctx| {
+            // The swipe loop: two drags toward newer content, two back --
+            // a cycle returns to where it started, so the whole loop
+            // measures steady-state scrolling. `BenchRun.kt`'s own
+            // comment on this shape.
+            for _ in 0..CYCLES {
+                for delta in [SWIPE_PX, SWIPE_PX, -SWIPE_PX, -SWIPE_PX] {
+                    animate_scroll(&mut ctx, &redraw, delta, SWIPE_MS).await;
+                    tokio::time::sleep(Duration::from_millis(SWIPE_PAUSE_MS)).await;
+                }
+            }
+
+            // Pinned to the newest end before streaming starts, matching
+            // `stream-bench.sh`'s "Jump to latest" tap.
+            ctx.update(|state: &mut BenchClient, rsc| {
+                if let Some(screen) = &state.screen {
+                    (screen.list)(rsc).jump_to_end();
+                }
+            });
+            redraw.request_redraw();
+
+            // The battery sampler runs concurrently with the streaming
+            // phase, once a second, the same cadence `BatterySampler` uses
+            // on the Compose side -- via its own JNI-attached thread, not
+            // `ctx.update`, since a sample needs no widget-tree access.
+            let sampler_done = Arc::new(AtomicBool::new(false));
+            let samples = Arc::new(std::sync::Mutex::new(Vec::<i32>::new()));
+            let sampler = platform.clone().map(|platform| {
+                let done = sampler_done.clone();
+                let samples = samples.clone();
+                tokio::spawn(async move {
+                    while !done.load(Ordering::Relaxed) {
+                        if let Some(value) = platform.battery_current_ua() {
+                            samples.lock().unwrap().push(value);
+                        }
+                        tokio::time::sleep(Duration::from_secs(1)).await;
+                    }
+                })
+            });
+
+            let total = (STREAM_EVENTS_PER_SEC * STREAM_SECONDS) as usize;
+            let mut sent = 0usize;
+            for event in stream_tail.into_iter().take(total) {
+                ctx.update(move |state: &mut BenchClient, rsc| {
+                    state.items = fold_event(&state.items, &event);
+                    state.rebuild_transcript(rsc);
+                });
+                redraw.request_redraw();
+                sent += 1;
+                tokio::time::sleep(Duration::from_millis(1000 / STREAM_EVENTS_PER_SEC)).await;
+            }
+            // Lets the last few deltas land and draw before the report is
+            // read -- `BenchRun.kt`'s own closing delay.
+            tokio::time::sleep(Duration::from_millis(300)).await;
+
+            sampler_done.store(true, Ordering::Relaxed);
+            if let Some(sampler) = sampler {
+                let _ = sampler.await;
+            }
+            let battery = battery_line(&samples.lock().unwrap());
+            let cpu_line = match (cpu_start, process_cpu_ms()) {
+                (Some(start), Some(end)) => {
+                    format!("  process CPU time over this run: {}ms", end.saturating_sub(start))
+                }
+                _ => "  process CPU time over this run: unavailable".to_string(),
+            };
+            let rss_line = match peak_rss_kb() {
+                Some(kb) => format!("  peak RSS: {kb}kB"),
+                None => "  peak RSS: unavailable (/proc/self/status unreadable)".to_string(),
+            };
+
+            ctx.update(move |state: &mut BenchClient, rsc| {
+                state.running = false;
+                let scroll_line = format!(
+                    "  scroll: {CYCLES} cycles ({} swipes), streamed {sent}/{total} fixture events",
+                    CYCLES * 4
+                );
+                let frames_line = match state.android_state().frame_report.report() {
+                    Some(stats) => format!("{stats}"),
+                    None => "no frames recorded".to_string(),
+                };
+                let report = format!(
+                    "iris bench report\n{frames_line}\n{scroll_line}\n{cpu_line}\n{rss_line}\n{battery}"
+                );
+                log::info!("iris bench report: {report}");
+                state.report_display.edit(rsc).set(&report);
+                state.last_report = Some(report);
+            });
+            redraw.request_redraw();
+        });
+    }
+}
+
+/// Moves `List::scroll` by `total_px` over `duration_ms`, in ~60Hz steps,
+/// so the swipe is many rendered frames rather than one jump -- the same
+/// shape `animateScrollBy(SWIPE_PX, tween(SWIPE_MS))` gives on the Compose
+/// side, in the one place the two backends have to differ (iris's `List`
+/// has no built-in tween, so this drives it by hand).
+async fn animate_scroll(
+    ctx: &mut iris::task::TaskCtx<Rsc>,
+    redraw: &Arc<dyn iris::task::RequestRedraw>,
+    total_px: f32,
+    duration_ms: u64,
+) {
+    let steps = (duration_ms / ANIM_STEP_MS).max(1);
+    let step_px = total_px / steps as f32;
+    for _ in 0..steps {
+        ctx.update(move |state: &mut BenchClient, rsc| {
+            if let Some(screen) = &state.screen {
+                (screen.list)(rsc).scroll(step_px);
+            }
+        });
+        redraw.request_redraw();
+        tokio::time::sleep(Duration::from_millis(ANIM_STEP_MS)).await;
+    }
+}
@@ -0,0 +1,134 @@
+//! JNI calls the `bench` feature needs that go through the shell's own
+//! Java side rather than anything `iris`/`android-view` already wraps:
+//! `BatteryManager.getIntProperty(BATTERY_PROPERTY_CURRENT_NOW)` for the
+//! per-second battery sample, and `ClipboardManager.setPrimaryClip` for
+//! the "Copy report" control (P0's iris half, docs/RUST.md). Neither is
+//! part of `android_view::context`'s own `Context`/`Resources` wrappers
+//! (that file's own `// TODO: more methods?`), so this calls them
+//! directly rather than growing that crate's wrapper for two one-off
+//! calls this crate alone needs.
+//!
+//! Holds its own `JavaVM` + `GlobalRef` to the view (handed in through
+//! [`iris::android::AndroidAppState::platform_ready`]) so it can attach
+//! whichever thread calls it -- the battery sampler runs on a background
+//! tokio task, not the UI thread the rest of `IrisViewPeer`'s JNI calls
+//! run on. `JavaVM::attach_current_thread` is safe to call from a thread
+//! already attached (the `jni` crate detects it and does not double
+//! attach), so no caller here needs to know or care which thread it is.
+
+use android_view::jni::{
+    JNIEnv, JavaVM,
+    objects::{GlobalRef, JObject, JValue},
+};
+
+/// `android.os.BatteryManager.BATTERY_PROPERTY_CURRENT_NOW` -- not exposed
+/// as a constant anywhere reachable without the Android SDK jar, so named
+/// here with its source rather than left as a bare `2`.
+const BATTERY_PROPERTY_CURRENT_NOW: i32 = 2;
+
+pub struct PlatformHandle {
+    vm: JavaVM,
+    view: GlobalRef,
+}
+
+impl PlatformHandle {
+    pub fn new(vm: JavaVM, view: GlobalRef) -> Self {
+        Self { vm, view }
+    }
+
+    fn context<'e>(&self, env: &mut JNIEnv<'e>) -> Option<JObject<'e>> {
+        env.call_method(
+            self.view.as_obj(),
+            "getContext",
+            "()Landroid/content/Context;",
+            &[],
+        )
+        .ok()?
+        .l()
+        .ok()
+    }
+
+    fn system_service<'e>(
+        &self,
+        env: &mut JNIEnv<'e>,
+        context: &JObject<'e>,
+        name: &str,
+    ) -> Option<JObject<'e>> {
+        let jname = env.new_string(name).ok()?;
+        env.call_method(
+            context,
+            "getSystemService",
+            "(Ljava/lang/String;)Ljava/lang/Object;",
+            &[JValue::Object(jname.as_ref())],
+        )
+        .ok()?
+        .l()
+        .ok()
+    }
+
+    /// One sample of `BATTERY_PROPERTY_CURRENT_NOW`, in microamps. `None`
+    /// on any JNI failure, on a device with no `BatteryManager` service,
+    /// or when the platform itself answers "not supported" -- `0` or
+    /// `Integer.MIN_VALUE` are both documented SDK answers for that, and
+    /// both would read as a real (and wrong) measurement if folded into an
+    /// average rather than named apart. UI_RULES.md: never present an
+    /// inferred value as a measured one.
+    pub fn battery_current_ua(&self) -> Option<i32> {
+        let mut guard = self.vm.attach_current_thread().ok()?;
+        let env: &mut JNIEnv = &mut guard;
+        let context = self.context(env)?;
+        let battery_manager = self.system_service(env, &context, "batterymanager")?;
+        let value = env
+            .call_method(
+                &battery_manager,
+                "getIntProperty",
+                "(I)I",
+                &[JValue::Int(BATTERY_PROPERTY_CURRENT_NOW)],
+            )
+            .ok()?
+            .i()
+            .ok()?;
+        if value == 0 || value == i32::MIN {
+            None
+        } else {
+            Some(value)
+        }
+    }
+
+    /// Puts `text` on the system clipboard through `ClipboardManager` --
+    /// `true` only if the whole JNI chain (service lookup, `ClipData`,
+    /// `setPrimaryClip`) succeeded.
+    pub fn copy_to_clipboard(&self, label: &str, text: &str) -> bool {
+        self.try_copy_to_clipboard(label, text).is_some()
+    }
+
+    fn try_copy_to_clipboard(&self, label: &str, text: &str) -> Option<()> {
+        let mut guard = self.vm.attach_current_thread().ok()?;
+        let env: &mut JNIEnv = &mut guard;
+        let context = self.context(env)?;
+        let clipboard = self.system_service(env, &context, "clipboard")?;
+        let jlabel = env.new_string(label).ok()?;
+        let jtext = env.new_string(text).ok()?;
+        let clip = env
+            .call_static_method(
+                "android/content/ClipData",
+                "newPlainText",
+                "(Ljava/lang/CharSequence;Ljava/lang/CharSequence;)Landroid/content/ClipData;",
+                &[
+                    JValue::Object(jlabel.as_ref()),
+                    JValue::Object(jtext.as_ref()),
+                ],
+            )
+            .ok()?
+            .l()
+            .ok()?;
+        env.call_method(
+            &clipboard,
+            "setPrimaryClip",
+            "(Landroid/content/ClipData;)V",
+            &[JValue::Object(&clip)],
+        )
+        .ok()?;
+        Some(())
+    }
+}
@@ -1,5 +1,5 @@
-//! The android-view demo app: iris's `tabs` widget tree (`tabs_ui::build`,
-//! shared with the winit example) running through
+//! The android-view demo app: by default, iris's `tabs` widget tree
+//! (`tabs_ui::build`, shared with the winit example) running through
 //! `iris::android`'s `ViewPeer`. This is RUST.md's I2 pass condition made
 //! concrete -- there is no UI here beyond what `tabs-ui` already draws.
 //!
@@ -9,6 +9,32 @@
 //! wrapping `iris::android::new_peer`'s generic function in a concrete
 //! `extern "system" fn`, since `register_view_class` wants a plain
 //! function pointer.
+//!
+//! **`transcript-screen` feature (RUST.md's I5 Android integration):** with
+//! `--features transcript-screen`, `new_view_peer` instantiates
+//! `transcript_client::TranscriptClient` instead of the tabs `Client`
+//! below, against a real `ai-server` (see that module's doc). Chosen over a
+//! third shell crate: this one already has the Gradle project, the
+//! `IrisView`/`MainActivity` Java, and the JNI registration I2 built and
+//! measured against, and the only thing a transcript screen needs on top
+//! is a different `AndroidAppState` -- the same axis `tabs_ui::build` vs.
+//! `transcript_ui::build` already varies along on the winit side (compare
+//! `iris/examples/tabs.rs` and `iris/transcript-ui/examples/transcript.rs`).
+//! A build picks one screen or the other, never both, so `Client` and
+//! `TranscriptClient` are cfg-gated apart rather than switched at runtime --
+//! there is no in-app navigation to switch *to* on either side yet.
+//!
+//! **`bench` feature (P0's iris half, docs/RUST.md):** a third
+//! `AndroidAppState`, `bench_client::BenchClient`, on the same axis --
+//! `transcript_ui::build_tree` again, this time against the checked-in
+//! fixture (`app/bench-fixture/assets/transcript.jsonl`) instead of a real
+//! server, with a "Run benchmark" control that drives the same scroll loop
+//! and streaming phase the Compose `bench` build type's `BenchRun.kt`
+//! does. `bench` depends on `transcript-screen` (Cargo.toml) for
+//! `transcript-ui`/`client-core`/`event-model`, so both features end up
+//! enabled together -- `ActiveClient` below gives `bench` priority in that
+//! case, the same way `transcript-screen` already takes priority over the
+//! default `tabs-screen`.

 use android_view::{
    Context, View,
@@ -18,19 +44,30 @@ use android_view::{
    },
    register_view_class,
 };
+#[cfg(not(feature = "transcript-screen"))]
 use iris::android::{AndroidAppState, AndroidRsc, AndroidUiState, HasAndroidUiState};
+#[cfg(not(feature = "transcript-screen"))]
 use iris::prelude::*;
 use log::LevelFilter;
 use std::ffi::c_void;

+#[cfg(feature = "bench")]
+mod bench_client;
+#[cfg(feature = "bench")]
+mod bench_jni;
+#[cfg(all(feature = "transcript-screen", not(feature = "bench")))]
+mod transcript_client;
+
 /// The app's `View` subclass, matching the Java side's package --
 /// `app/src/main/java/dev/iris/android/demo/IrisView.java`.
 const VIEW_CLASS: &str = "dev/iris/android/demo/IrisView";

+#[cfg(not(feature = "transcript-screen"))]
 pub struct Client {
    ui_state: AndroidUiState,
 }

+#[cfg(not(feature = "transcript-screen"))]
 impl HasAndroidUiState for Client {
    fn android_state(&self) -> &AndroidUiState {
        &self.ui_state
@@ -40,6 +77,7 @@ impl HasAndroidUiState for Client {
    }
 }

+#[cfg(not(feature = "transcript-screen"))]
 impl AndroidAppState for Client {
    fn new(mut ui_state: AndroidUiState, rsc: &mut AndroidRsc<Self>) -> Self {
        // `widgets.info` is the winit example's frame-debug readout, kept
@@ -61,12 +99,19 @@ impl AndroidAppState for Client {
    }
 }

+#[cfg(not(feature = "transcript-screen"))]
+type ActiveClient = Client;
+#[cfg(all(feature = "transcript-screen", not(feature = "bench")))]
+type ActiveClient = transcript_client::TranscriptClient;
+#[cfg(feature = "bench")]
+type ActiveClient = bench_client::BenchClient;
+
 extern "system" fn new_view_peer<'local>(
    env: JNIEnv<'local>,
    view: View<'local>,
    context: Context<'local>,
 ) -> jlong {
-    iris::android::new_peer::<Client>(env, view, context)
+    iris::android::new_peer::<ActiveClient>(env, view, context)
 }

 /// # Safety
@@ -0,0 +1,375 @@
+//! RUST.md's I5 Android integration: `transcript-ui`'s screen filling the
+//! whole window on android-view, against a real `ai-server` through
+//! `client-core` -- the missing half `iris-android-app` (I2) only had for
+//! `tabs-ui` until now. Behind the `transcript-screen` Cargo feature so the
+//! plain build (`cargo ndk build`, no `--features`) stays exactly the tabs
+//! demo I2/I4 already measured against.
+//!
+//! **Deliberate simplification, recorded rather than left to be
+//! rediscovered (RUST.md's I5 box has the full account)**: there is no
+//! session list and no enrollment UI here. The server, port, token and
+//! pinned CA are baked in at build time (`build.rs`'s
+//! `AI_APP_TRANSCRIPT_HOST`/`_PORT`/`_TOKEN`/`AI_APP_CA`), and the first
+//! session `ApiClient::fetch_sessions` returns is opened automatically --
+//! there is nothing to tap to get there, which is what `transcript-bench.sh`
+//! and `ui-trace` need to land straight on the screen under test. A real
+//! app needs `desktop-app`'s `EnrolledServer`/QR-link flow or E3's
+//! Keystore-sealed `ServerConfig.kt`; building a second one of those was
+//! not this pass's job.
+//!
+//! **Reuses `iris/desktop-app`'s `app.rs` shape almost exactly** --
+//! `fold_event`/`group_tool_runs`/`fold_page`/`raw_seq` from
+//! `client_core::transcript_fold`, a `generation` counter guarding against
+//! a stale background response, and a full rebuild of the widget tree on
+//! every event (same tradeoff, same reason: `push_row` cannot update a row
+//! already on screen, and this rig's conversations are small). What
+//! differs is only the redraw mechanism: android-view has no
+//! `winit::EventLoopProxy`, so this uses `iris::task::Tasks::redraw_handle`
+//! (new, added alongside this box) to request a frame after each
+//! `TaskCtx::update` instead of relying on `Tasks::spawn`'s single
+//! end-of-future redraw -- see that method's own doc for why.
+
+use client_core::api::{ApiClient, UreqTransport};
+use client_core::event_stream::{StreamItem, follow_session_events};
+use client_core::transcript_fold::{TranscriptItem, fold_event, fold_page, group_tool_runs};
+use event_model::SeqEvent;
+use iris::android::{AndroidAppState, AndroidRsc, AndroidUiState, HasAndroidUiState};
+use iris::prelude::*;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicU64, Ordering};
+
+mod pinned {
+    include!(concat!(env!("OUT_DIR"), "/pinned_config.rs"));
+}
+
+pub struct TranscriptClient {
+    ui_state: AndroidUiState,
+    /// The screen's own content -- everything under the fixed
+    /// [`frame_report_controls`] bar, which is built once (`new`, below)
+    /// and never touched by `show_message`/`rebuild_transcript`'s own
+    /// `set` calls the way `desktop-app`'s `transcript_ptr` isn't touched
+    /// by rebuilding the session list beside it.
+    content: WeakWidget<WidgetPtr>,
+    screen: Option<transcript_ui::TranscriptScreen>,
+    /// The folded transcript as of the last rebuild -- kept here (not
+    /// re-derived) for the same reason `desktop-app`'s `Client::items`
+    /// exists: a live `StreamEvent` only carries one new wire event, and
+    /// `fold_event` needs everything folded so far to fold it in.
+    items: Vec<TranscriptItem>,
+    /// The session currently open -- `None` only before the first fetch
+    /// resolves. Read back by `apply_event`'s rebuild, which has no session
+    /// id of its own (a live `SeqEvent` doesn't carry one).
+    session_id: Option<String>,
+    /// Bumped every time a new session load starts; a background response
+    /// checks it before touching state, so a slow reply for a session this
+    /// screen has moved on from can't overwrite what replaced it. There is
+    /// only ever one session here (no list to switch away to), but the
+    /// guard still matters for the *first* fetch racing a `stop`/`start`.
+    generation: Arc<AtomicU64>,
+}
+
+impl HasAndroidUiState for TranscriptClient {
+    fn android_state(&self) -> &AndroidUiState {
+        &self.ui_state
+    }
+    fn android_state_mut(&mut self) -> &mut AndroidUiState {
+        &mut self.ui_state
+    }
+}
+
+/// Builds one `UreqTransport` from the config `build.rs` baked in. Called
+/// twice per session load, same as `desktop-app`'s `build_transport`
+/// closure -- `ApiClient` and the live-stream follow each need their own,
+/// since `UreqTransport` holds its own `ureq::Agent`.
+fn build_transport() -> Result<UreqTransport, String> {
+    let base_url = format!("https://{}:{}", pinned::HOST, pinned::PORT);
+    UreqTransport::new(
+        base_url,
+        pinned::TOKEN.to_string(),
+        pinned::CA_PEM.as_bytes(),
+    )
+    .map_err(|e| e.to_string())
+}
+
+fn placeholder<Rsc: HasEvents>(rsc: &mut Rsc, message: &str) -> StrongWidget {
+    wtext(message.to_string())
+        .color(Color::WHITE)
+        .wrap(true)
+        .pad(16)
+        .add_strong(rsc)
+        .any()
+}
+
+/// The two named controls RUST.md's I5 box ("Measurements taken" (b))
+/// drives by name over `ui-trace`, e.g. `ui-trace record --do "tap 'Frame
+/// report'"`. `dumpsys gfxinfo` cannot see this screen's own GPU-drawn
+/// frames at all -- this is the screen's own equivalent of the Compose
+/// app's "Copy render timings" control, logged rather than clipboarded
+/// (no clipboard wiring exists here) under this crate's own fixed
+/// `android_logger` tag (`iris-android-app`, `lib.rs`'s `JNI_OnLoad`),
+/// grep-able on the fixed string `"iris frame report"` the way
+/// `transcript-bench.sh` greps `"ai-app render report"`.
+fn frame_report_controls(rsc: &mut AndroidRsc<TranscriptClient>) -> WeakWidget {
+    type Rsc = AndroidRsc<TranscriptClient>;
+    let report_rect = rect(Color::rgb(50, 50, 60))
+        .on(
+            CursorSense::click(),
+            |ctx: EventIdCtx<'_, Rsc, _, _>, _rsc: &mut Rsc| match ctx
+                .state
+                .android_state()
+                .frame_report
+                .report()
+            {
+                Some(stats) => log::info!("iris frame report: {stats}"),
+                None => log::info!(
+                    "iris frame report: no frames recorded -- scroll first, then press this"
+                ),
+            },
+        )
+        .label("Frame report");
+    let report = (
+        report_rect,
+        wtext("Frame report").size(18).text_align(Align::CENTER),
+    )
+        .stack()
+        .pad(8)
+        .add(rsc);
+
+    let reset_rect = rect(Color::rgb(70, 40, 40))
+        .on(
+            CursorSense::click(),
+            |ctx: EventIdCtx<'_, Rsc, _, _>, _rsc: &mut Rsc| {
+                ctx.state.android_state_mut().frame_report.reset();
+                log::info!("iris frame report: reset");
+            },
+        )
+        .label("Reset frame report");
+    let reset = (
+        reset_rect,
+        wtext("Reset").size(18).text_align(Align::CENTER),
+    )
+        .stack()
+        .pad(8)
+        .add(rsc);
+
+    (report, reset).span(Dir::RIGHT).height(56).add(rsc)
+}
+
+impl AndroidAppState for TranscriptClient {
+    fn new(mut ui_state: AndroidUiState, rsc: &mut AndroidRsc<Self>) -> Self {
+        let content = WidgetPtr::new().add(rsc);
+        let loading = placeholder(rsc, "Loading sessions...");
+        content(rsc).set(loading);
+
+        let tree = (frame_report_controls(rsc), content.height(rest(1)))
+            .span(Dir::DOWN)
+            .add_strong(rsc)
+            .any();
+        ui_state.set_root(tree);
+
+        let mut client = Self {
+            ui_state,
+            content,
+            screen: None,
+            items: Vec::new(),
+            session_id: None,
+            generation: Arc::new(AtomicU64::new(0)),
+        };
+        client.spawn_fetch_sessions(rsc);
+        client
+    }
+
+    fn back_pressed(&mut self, _rsc: &mut AndroidRsc<Self>, _render: &mut UiRenderState) -> bool {
+        // No screen stack of its own -- same "let the activity finish"
+        // answer `iris-android-app`'s tabs `Client` already gives.
+        false
+    }
+}
+
+impl TranscriptClient {
+    fn show_message(&mut self, rsc: &mut AndroidRsc<Self>, message: &str) {
+        let widget = placeholder(rsc, message);
+        (self.content)(rsc).set(widget);
+        self.screen = None;
+    }
+
+    fn spawn_fetch_sessions(&mut self, rsc: &mut AndroidRsc<Self>) {
+        let redraw = rsc.tasks.redraw_handle();
+        let my_generation = self.generation.load(Ordering::SeqCst);
+        let generation = self.generation.clone();
+        rsc.spawn_task(async move |mut ctx| {
+            let outcome = match build_transport() {
+                Ok(transport) => ApiClient::new(transport)
+                    .fetch_sessions()
+                    .map_err(|e| e.to_string()),
+                Err(e) => Err(format!("couldn't set up TLS: {e}")),
+            };
+            ctx.update(move |state: &mut TranscriptClient, rsc| {
+                if generation.load(Ordering::SeqCst) != my_generation {
+                    return;
+                }
+                match outcome {
+                    Ok(sessions) => match sessions.into_iter().next() {
+                        Some(session) => state.select_session(rsc, session.id),
+                        None => state.show_message(rsc, "No sessions on the sandbox server."),
+                    },
+                    Err(message) => {
+                        state.show_message(rsc, &format!("Couldn't list sessions: {message}"))
+                    }
+                }
+            });
+            redraw.request_redraw();
+        });
+    }
+
+    /// Loads the opening page, then follows the live SSE stream for the
+    /// rest of this session's life -- `desktop-app`'s `select_session`
+    /// almost verbatim, with `Proxy::send_event` replaced by `ctx.update` +
+    /// `redraw.request_redraw()` (see this module's doc).
+    fn select_session(&mut self, rsc: &mut AndroidRsc<Self>, session_id: String) {
+        let my_generation = self.generation.fetch_add(1, Ordering::SeqCst) + 1;
+        self.items.clear();
+        self.session_id = Some(session_id.clone());
+        self.show_message(rsc, "Loading transcript...");
+
+        let redraw = rsc.tasks.redraw_handle();
+        let live_generation = self.generation.clone();
+        rsc.spawn_task(async move |mut ctx| {
+            let transports =
+                build_transport().and_then(|rest| build_transport().map(|stream| (rest, stream)));
+            let (rest, stream_transport) = match transports {
+                Ok(pair) => pair,
+                Err(e) => {
+                    let message = format!("couldn't set up TLS: {e}");
+                    ctx.update(move |state: &mut TranscriptClient, rsc| {
+                        if live_generation.load(Ordering::SeqCst) == my_generation {
+                            state.show_message(rsc, &message);
+                        }
+                    });
+                    redraw.request_redraw();
+                    return;
+                }
+            };
+            let api = ApiClient::new(rest);
+
+            // The most recent 200 events, coalesced -- the same page size
+            // `desktop-app` uses; RUST.md's I3/history-paging work is what
+            // a real scrollback would reuse (out of scope here, same as
+            // E4).
+            let page: Result<Vec<serde_json::Value>, String> = api
+                .fetch_transcript_page(&session_id, None, 200, true)
+                .map_err(|e| e.to_string());
+            // The wire `seq` of the last line, not a folded item's `seq()`
+            // -- see `client_core::transcript_fold::raw_seq`'s doc for why
+            // resuming from the latter re-delivers deltas already folded
+            // into an in-progress reply.
+            let after = page
+                .as_ref()
+                .ok()
+                .and_then(|values| values.last())
+                .and_then(client_core::transcript_fold::raw_seq)
+                .unwrap_or(0);
+            let result = page.and_then(|values| fold_page(&values));
+
+            {
+                let live_generation = live_generation.clone();
+                ctx.update(move |state: &mut TranscriptClient, rsc| {
+                    if live_generation.load(Ordering::SeqCst) != my_generation {
+                        return;
+                    }
+                    match result {
+                        Ok(items) => {
+                            state.items = items;
+                            state.rebuild_transcript(rsc);
+                        }
+                        Err(message) => {
+                            state.show_message(rsc, &format!("Couldn't load transcript: {message}"))
+                        }
+                    }
+                });
+            }
+            redraw.request_redraw();
+
+            if live_generation.load(Ordering::SeqCst) != my_generation {
+                return;
+            }
+            // The outer closure here is an `FnMut` -- `follow_session_events`
+            // calls it once per line -- so it captures `live_generation` by
+            // move and re-clones it for each inner `ctx.update` closure
+            // rather than moving a shared `stop`-style helper into itself:
+            // a value moved out of an `FnMut`'s captures on one call leaves
+            // nothing there for the next.
+            let _ =
+                follow_session_events(
+                    &stream_transport,
+                    &session_id,
+                    after,
+                    move |item| match item {
+                        StreamItem::Open | StreamItem::Reset => {
+                            live_generation.load(Ordering::SeqCst) == my_generation
+                        }
+                        StreamItem::Event { event, .. } => {
+                            if live_generation.load(Ordering::SeqCst) != my_generation {
+                                return false;
+                            }
+                            let live_generation = live_generation.clone();
+                            ctx.update(move |state: &mut TranscriptClient, rsc| {
+                                if live_generation.load(Ordering::SeqCst) != my_generation {
+                                    return;
+                                }
+                                state.apply_event(rsc, &event);
+                            });
+                            redraw.request_redraw();
+                            true
+                        }
+                    },
+                );
+        });
+    }
+
+    /// Rebuilds the whole widget tree from `self.items` -- same tradeoff as
+    /// `desktop-app`'s `rebuild_transcript` (this module's doc comment).
+    /// Reads `self.session_id` rather than taking one, since every caller
+    /// (the opening page, and every live event) already has it set there.
+    fn rebuild_transcript(&mut self, rsc: &mut AndroidRsc<Self>) {
+        let in_progress = self
+            .screen
+            .as_ref()
+            .map(|screen| screen.composer.field.edit(rsc).text.text().to_string())
+            .filter(|t| !t.is_empty());
+
+        let rows = group_tool_runs(&self.items);
+        let (screen, tree) = transcript_ui::build_tree(rsc, rows);
+
+        if let Some(text) = in_progress {
+            screen.composer.field.edit(rsc).set(&text);
+        }
+        if let Some(session_id) = self.session_id.clone() {
+            let field = screen.composer.field;
+            rsc.register_event(field, Submit, move |ctx, rsc| {
+                let text = field.edit(rsc).take();
+                let text = text.trim().to_string();
+                if !text.is_empty() {
+                    ctx.state.send_message(session_id.clone(), text);
+                }
+            });
+        }
+
+        (self.content)(rsc).set(tree);
+        self.screen = Some(screen);
+    }
+
+    fn apply_event(&mut self, rsc: &mut AndroidRsc<Self>, event: &SeqEvent) {
+        self.items = fold_event(&self.items, event);
+        self.rebuild_transcript(rsc);
+    }
+
+    fn send_message(&mut self, session_id: String, text: String) {
+        std::thread::spawn(move || {
+            if let Ok(transport) = build_transport() {
+                let api = ApiClient::new(transport);
+                let _ = api.send_message(&session_id, &text, &[]);
+            }
+        });
+    }
+}
@@ -0,0 +1,299 @@
+use std::time::Duration;
+
+/// The frame budget `dumpsys gfxinfo` also uses to call a frame "janky": the
+/// 60Hz vsync period. Kept as the same threshold so a percentage from this
+/// report and a percentage from `gfxinfo` mean the same thing.
+pub const JANK_THRESHOLD: Duration = Duration::from_nanos(16_666_667);
+
+/// Enough frames for several minutes of scrolling before the oldest ones
+/// start being overwritten -- the same "diagnostic, not a log" sizing
+/// `FrameStats.kt`'s `CAP` uses on the Compose side, chosen independently
+/// here since a `Duration` is smaller than the six `Long` arrays it keeps.
+const RING_CAPACITY: usize = 4096;
+
+/// A per-frame wall-time report iris keeps of itself, because `dumpsys
+/// gfxinfo` cannot see a `SurfaceView`'s own GPU-drawn frames at all
+/// (RUST.md's I5 box, "Measurements taken" (b)): it instruments Android's
+/// ordinary Skia/HWUI View-drawing pipeline, which a `wgpu`-rendered
+/// `SurfaceView` bypasses entirely. `record` is meant to be called once per
+/// frame, wrapping the same span Compose's own render report and `gfxinfo`
+/// count -- from the frame's redraw/update start to after the frame is
+/// handed to the platform to present.
+///
+/// **What this does not measure**: wgpu's `present()` call queues the frame
+/// with the compositor and returns; it is not fenced against the GPU
+/// actually finishing the frame or the compositor actually showing it, the
+/// way `gfxinfo`'s own `GPU_DURATION`/vsync accounting is. So a sample here
+/// is "how long the CPU took to build and submit this frame", not
+/// "how long the frame took to reach the screen" -- named in
+/// [`FrameStats`]'s own `Display` line rather than presented as the latter,
+/// per the standing rule against showing an inferred number as a measured
+/// one where the two differ.
+///
+/// Fixed-size ring, no allocation on the hot path -- `report()` is the only
+/// place that allocates (a sort over the current ring), and it is only
+/// ever called from a button tap, not once per frame.
+pub struct FrameReport {
+    ring: Box<[Duration; RING_CAPACITY]>,
+    /// The `submit_to_present` half of each sample in `ring`, same index,
+    /// same lifetime -- kept as a second ring rather than a ring of pairs so
+    /// the existing `ring`/percentile code above is untouched (RUST.md's I5
+    /// "Where iris's frame time goes" CPU/GPU split, added 2026-09-05).
+    /// `ring[i] - submit_ring[i]` is that frame's `redraw_to_submit` half.
+    submit_ring: Box<[Duration; RING_CAPACITY]>,
+    /// How many of `ring`'s slots hold a real sample -- saturates at
+    /// `RING_CAPACITY`, unlike `total_frames` below which keeps counting.
+    len: usize,
+    pos: usize,
+    /// All frames recorded since the last `reset`, even past `RING_CAPACITY`
+    /// -- what `janky_percent` divides by, so a long run's percentage stays
+    /// correct even once the ring itself only holds the most recent frames.
+    total_frames: u64,
+    janky_frames: u64,
+}
+
+/// One resolved reading. `Display` is the log line both the "Frame report"
+/// button and `transcript-bench.sh`-style scripts read, grep-able on
+/// `"iris frame report"`.
+pub struct FrameStats {
+    pub total_frames: u64,
+    pub janky_percent: f64,
+    pub p50: Duration,
+    pub p90: Duration,
+    pub p99: Duration,
+    pub worst: Duration,
+    /// Median of `redraw_to_submit` -- iris's own CPU work (layout, text,
+    /// primitive building) up to and including building the `queue.submit`
+    /// call, per frame. RUST.md's I5 "Where iris's frame time goes" split,
+    /// added 2026-09-05 to answer "CPU or GPU?" with a number rather than a
+    /// guess.
+    pub cpu_p50: Duration,
+    /// Median of `submit_to_present` -- the `queue.submit` call itself plus
+    /// `present()`, i.e. wherever the driver/GPU/compositor wait actually
+    /// happens. Same caveat as the type's own doc: `present()` is not
+    /// fenced against the GPU actually finishing, so this is "how long the
+    /// CPU was blocked handing the frame off", not the frame's true GPU
+    /// time -- still enough to separate "iris is slow building the frame"
+    /// from "iris is slow handing it to the driver".
+    pub gpu_wait_p50: Duration,
+}
+
+impl std::fmt::Display for FrameStats {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(
+            f,
+            "frames={} janky%={:.2} p50={:.1}ms p90={:.1}ms p99={:.1}ms worst={:.1}ms \
+             (measures redraw-start to after present() is called, not GPU/compositor \
+             completion)",
+            self.total_frames,
+            self.janky_percent,
+            self.p50.as_secs_f64() * 1000.0,
+            self.p90.as_secs_f64() * 1000.0,
+            self.p99.as_secs_f64() * 1000.0,
+            self.worst.as_secs_f64() * 1000.0,
+        )?;
+        write!(
+            f,
+            " cpu_p50={:.1}ms gpu_wait_p50={:.1}ms (redraw-start-to-submit vs. \
+             submit-to-after-present)",
+            self.cpu_p50.as_secs_f64() * 1000.0,
+            self.gpu_wait_p50.as_secs_f64() * 1000.0,
+        )
+    }
+}
+
+impl FrameReport {
+    pub fn new() -> Self {
+        Self {
+            ring: Box::new([Duration::ZERO; RING_CAPACITY]),
+            submit_ring: Box::new([Duration::ZERO; RING_CAPACITY]),
+            len: 0,
+            pos: 0,
+            total_frames: 0,
+            janky_frames: 0,
+        }
+    }
+
+    /// Record one frame's elapsed wall time, with no CPU/GPU split (the
+    /// `submit_to_present` half is recorded as zero, so `cpu_p50` reads as
+    /// the whole frame and `gpu_wait_p50` as nothing -- honest for a caller
+    /// that never measured the split, rather than fabricating one). O(1),
+    /// no allocation.
+    pub fn record(&mut self, elapsed: Duration) {
+        self.record_split(elapsed, Duration::ZERO);
+    }
+
+    /// Record one frame's elapsed wall time, split at `queue.submit`:
+    /// `submit_to_present` is the `queue.submit()` call plus `present()`;
+    /// `total - submit_to_present` is everything before it (layout, text,
+    /// primitive building). RUST.md's I5 "Where iris's frame time goes"
+    /// CPU/GPU split, added 2026-09-05. O(1), no allocation.
+    pub fn record_split(&mut self, total: Duration, submit_to_present: Duration) {
+        self.ring[self.pos] = total;
+        self.submit_ring[self.pos] = submit_to_present;
+        self.pos = (self.pos + 1) % RING_CAPACITY;
+        self.len = (self.len + 1).min(RING_CAPACITY);
+        self.total_frames += 1;
+        if total > JANK_THRESHOLD {
+            self.janky_frames += 1;
+        }
+    }
+
+    /// Clears every counter and every sample -- what the "Reset frame
+    /// report" control calls, so a report covers only what was scrolled
+    /// after the button was pressed (the same reason `FrameStats.kt`'s
+    /// `reset()` exists on the Compose side).
+    pub fn reset(&mut self) {
+        self.len = 0;
+        self.pos = 0;
+        self.total_frames = 0;
+        self.janky_frames = 0;
+    }
+
+    /// `None` if nothing has been recorded since the last reset -- the
+    /// "no frames recorded, scroll first" case, not a zeroed report that
+    /// would read as a real (perfect) measurement.
+    pub fn report(&self) -> Option<FrameStats> {
+        if self.len == 0 {
+            return None;
+        }
+        let mut samples: Vec<Duration> = self.ring[..self.len].to_vec();
+        samples.sort_unstable();
+        let pct = |p: usize| samples[(samples.len() * p / 100).min(samples.len() - 1)];
+
+        // Separate arrays rather than subtracting the two medians above:
+        // medians do not distribute over subtraction, and each needs its
+        // own sort.
+        let submit_samples: Vec<Duration> = self.submit_ring[..self.len].to_vec();
+        let cpu_samples: Vec<Duration> = self.ring[..self.len]
+            .iter()
+            .zip(self.submit_ring[..self.len].iter())
+            .map(|(&total, &submit_to_present)| total.saturating_sub(submit_to_present))
+            .collect();
+        let median = |mut v: Vec<Duration>| {
+            v.sort_unstable();
+            v[v.len() / 2]
+        };
+
+        Some(FrameStats {
+            total_frames: self.total_frames,
+            janky_percent: 100.0 * self.janky_frames as f64 / self.total_frames as f64,
+            p50: pct(50),
+            p90: pct(90),
+            p99: pct(99),
+            worst: *samples.last().expect("len > 0 checked above"),
+            cpu_p50: median(cpu_samples),
+            gpu_wait_p50: median(submit_samples),
+        })
+    }
+}
+
+impl Default for FrameReport {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn no_frames_reports_none() {
+        assert!(FrameReport::new().report().is_none());
+    }
+
+    #[test]
+    fn one_frame_is_every_percentile_and_the_worst() {
+        let mut r = FrameReport::new();
+        r.record(Duration::from_millis(10));
+        let stats = r.report().unwrap();
+        assert_eq!(stats.total_frames, 1);
+        assert_eq!(stats.p50, Duration::from_millis(10));
+        assert_eq!(stats.p99, Duration::from_millis(10));
+        assert_eq!(stats.worst, Duration::from_millis(10));
+        assert_eq!(stats.janky_percent, 0.0);
+    }
+
+    #[test]
+    fn percentiles_and_worst_over_a_known_set() {
+        let mut r = FrameReport::new();
+        // 100 samples, 1ms..=100ms, fed out of order so the ring's own
+        // order is not what gives the right answer -- the sort has to.
+        for ms in (1..=100).rev() {
+            r.record(Duration::from_millis(ms));
+        }
+        let stats = r.report().unwrap();
+        assert_eq!(stats.total_frames, 100);
+        assert_eq!(stats.p50, Duration::from_millis(51));
+        assert_eq!(stats.p90, Duration::from_millis(91));
+        assert_eq!(stats.p99, Duration::from_millis(100));
+        assert_eq!(stats.worst, Duration::from_millis(100));
+    }
+
+    #[test]
+    fn jank_threshold_matches_gfxinfos_60hz_budget() {
+        let mut r = FrameReport::new();
+        r.record(Duration::from_nanos(16_666_667)); // exactly on budget: not janky
+        r.record(Duration::from_nanos(16_666_668)); // one ns over: janky
+        let stats = r.report().unwrap();
+        assert_eq!(stats.janky_percent, 50.0);
+    }
+
+    #[test]
+    fn janky_percent_is_over_all_time_frames_not_just_the_ring() {
+        // Fewer than RING_CAPACITY frames, all janky, then a fresh reset --
+        // the percentage must reset to 0, not divide by a stale count.
+        let mut r = FrameReport::new();
+        for _ in 0..10 {
+            r.record(Duration::from_millis(50));
+        }
+        assert_eq!(r.report().unwrap().janky_percent, 100.0);
+        r.reset();
+        assert!(r.report().is_none());
+        r.record(Duration::from_millis(1));
+        assert_eq!(r.report().unwrap().janky_percent, 0.0);
+    }
+
+    #[test]
+    fn record_without_a_split_reports_the_whole_frame_as_cpu() {
+        // A caller that never measured the split (plain `record`) should
+        // not fabricate a GPU-wait number -- it reads as zero, and the CPU
+        // half reads as the whole frame.
+        let mut r = FrameReport::new();
+        r.record(Duration::from_millis(20));
+        let stats = r.report().unwrap();
+        assert_eq!(stats.cpu_p50, Duration::from_millis(20));
+        assert_eq!(stats.gpu_wait_p50, Duration::ZERO);
+    }
+
+    #[test]
+    fn record_split_reports_each_halfs_own_median() {
+        let mut r = FrameReport::new();
+        // Three frames: total is always 30ms, but the CPU/GPU-wait split
+        // moves, so the two medians must be independent of each other and
+        // of `total`'s own median.
+        r.record_split(Duration::from_millis(30), Duration::from_millis(5));
+        r.record_split(Duration::from_millis(30), Duration::from_millis(10));
+        r.record_split(Duration::from_millis(30), Duration::from_millis(20));
+        let stats = r.report().unwrap();
+        assert_eq!(stats.p50, Duration::from_millis(30));
+        assert_eq!(stats.gpu_wait_p50, Duration::from_millis(10));
+        assert_eq!(stats.cpu_p50, Duration::from_millis(20));
+    }
+
+    #[test]
+    fn ring_wraps_without_growing_past_capacity() {
+        let mut r = FrameReport::new();
+        for i in 0..(RING_CAPACITY * 2) {
+            r.record(Duration::from_millis(1 + (i % 5) as u64));
+        }
+        let stats = r.report().unwrap();
+        // total_frames keeps the full count even once the ring has wrapped.
+        assert_eq!(stats.total_frames, (RING_CAPACITY * 2) as u64);
+        // but every sample the ring can report on is still one of the five
+        // values fed in, since a wrap can only overwrite with more of the
+        // same pattern here.
+        assert!(stats.worst <= Duration::from_millis(5));
+    }
+}
@@ -11,16 +11,60 @@ use wgpu::{

 mod atlas;
 mod data;
+mod frame_report;
 mod primitive;
 mod texture;
 mod util;

 pub use atlas::*;
 pub use data::{Mask, MaskIdx, MoveIdx, MoveOffset};
+pub use frame_report::{FrameReport, FrameStats, JANK_THRESHOLD};
 pub use primitive::*;

 const SHAPE_SHADER: &str = include_str!("./shader.wgsl");

+/// The `wgpu::Limits` both platform backends (`android::render::
+/// AndroidRenderer::new`, `default::render::UiRenderer::new`) ask
+/// `Adapter::request_device` for -- shared so the two copies cannot drift,
+/// per AGENTS.md's "write the logic once."
+///
+/// Built from `Limits::default()`, **not** a downlevel variant: the shader
+/// (`shader.wgsl`) reads four `var<storage>` buffers (rects, glyphs, masks,
+/// move_offsets) from the vertex stage, and `Limits::downlevel_webgl2_defaults()`
+/// zeroes `max_storage_buffers_per_shader_stage` along with the compute
+/// limits below -- switching to it would trade one `request_device` crash
+/// for a bind-group-layout one on the same downlevel hardware this is meant
+/// to support. `max_buffer_size` is raised for the growing instance/atlas
+/// buffers (`ArrBuf`, `GpuTextures`); everything else is `default()`'s
+/// desktop-tier value, unchanged.
+///
+/// The six `max_compute_*` fields are zeroed because nothing in this crate
+/// creates a `ComputePipeline` or writes a `@compute` shader stage --
+/// grepped for both across `iris`/`iris-core` before writing this, found
+/// none. `Limits::default()` requests desktop-tier compute limits
+/// unconditionally (`max_compute_workgroups_per_dimension: 65535`) even
+/// though nothing asks a device to actually support compute, which is what
+/// crashed `request_device` on the Android emulator's software GL path
+/// (`EMU_GPU=software`, `force-gles`): SwiftShader's GL reports itself as
+/// OpenGL ES 3.0, which has no compute shaders, so the adapter's real limit
+/// is 0 and the unconditional request fails outright
+/// (`RUST.md`'s "Software mode ... crashes for a third, different reason").
+/// The same would happen on a real GLES-3.0-only Android device. If a
+/// future change adds a compute pass, request the specific limits it needs
+/// here rather than reverting to the desktop-tier default for everything.
+pub fn device_limits() -> Limits {
+    Limits {
+        max_buffer_size: 1 << 30,
+        max_compute_workgroup_storage_size: 0,
+        max_compute_invocations_per_workgroup: 0,
+        max_compute_workgroup_size_x: 0,
+        max_compute_workgroup_size_y: 0,
+        max_compute_workgroup_size_z: 0,
+        max_compute_workgroups_per_dimension: 0,
+        ..Default::default()
+    }
+}
+
 pub struct UiRenderNode {
    uniform_group: BindGroup,
    primitive_layout: BindGroupLayout,
@@ -45,7 +45,9 @@

 use client_core::api::{ApiClient, SessionSummary, UreqTransport};
 use client_core::event_stream::{StreamItem, follow_session_events};
-use client_core::transcript_fold::{TranscriptItem, fold_event, group_tool_runs};
+use client_core::transcript_fold::{
+    TranscriptItem, fold_event, fold_page, group_tool_runs, raw_seq,
+};
 use event_model::SeqEvent;
 use iris::prelude::*;
 use std::sync::Arc;
@@ -421,118 +423,3 @@ fn placeholder(rsc: &mut DefaultRsc<Client>, message: &str) -> StrongWidget {
        .add_strong(rsc)
        .any()
 }
-
-/// Folds a page of raw transcript lines (`ApiClient::fetch_transcript_page`'s
-/// `Vec<Value>`) into the flat item list `client_core::transcript_fold`
-/// works over. A line this build can't parse fails the whole page rather
-/// than being skipped -- CODE_RULES's "an enumeration must be able to say
-/// 'it broke'" -- since silently dropping one event could hide, say, the
-/// user message the composer is about to look like it never sent.
-fn fold_page(values: &[serde_json::Value]) -> Result<Vec<TranscriptItem>, String> {
-    let mut items = Vec::new();
-    for value in values {
-        let event: SeqEvent = serde_json::from_value(value.clone()).map_err(|e| {
-            format!("the server sent a transcript line this build couldn't parse: {e}")
-        })?;
-        items = fold_event(&items, &event);
-    }
-    Ok(items)
-}
-
-/// The wire `seq` a raw transcript line carries -- see `select_session`'s
-/// comment on why the live-stream cursor has to be this, not a folded
-/// item's `seq()`.
-fn raw_seq(value: &serde_json::Value) -> Option<u64> {
-    value.get("seq")?.as_u64()
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    fn line(seq: u64, json: serde_json::Value) -> serde_json::Value {
-        let mut obj = json;
-        obj["seq"] = serde_json::json!(seq);
-        obj["ts"] = serde_json::json!(1.0);
-        obj
-    }
-
-    /// The regression for the bug a real `run-headless.sh` screenshot
-    /// found (see `select_session`'s comment): resuming the live stream
-    /// from the last *item's* seq re-delivers the deltas already folded
-    /// into a still-open assistant message, doubling its tail. `raw_seq`
-    /// of the last wire line must be the true high-water mark instead,
-    /// which for a run of deltas is higher than every item's own `seq()`.
-    #[test]
-    fn the_resume_cursor_is_the_last_wire_seq_not_the_last_items_seq() {
-        let values = vec![
-            line(1, serde_json::json!({"type": "userMessage", "text": "hi"})),
-            line(
-                2,
-                serde_json::json!({"type": "assistantText", "delta": "a"}),
-            ),
-            line(
-                3,
-                serde_json::json!({"type": "assistantText", "delta": "b"}),
-            ),
-            line(
-                4,
-                serde_json::json!({"type": "assistantText", "delta": "c"}),
-            ),
-        ];
-        let after = raw_seq(values.last().unwrap()).unwrap();
-        assert_eq!(after, 4);
-
-        let items = fold_page(&values).unwrap();
-        // The folded item keeps the *first* delta's seq (2), which is
-        // exactly the value that must not be used as the resume cursor.
-        let assistant_seq = items
-            .iter()
-            .find(|i| matches!(i, TranscriptItem::AssistantMsg { .. }))
-            .unwrap()
-            .seq();
-        assert_eq!(assistant_seq, 2);
-        assert_ne!(
-            after, assistant_seq,
-            "the fixed bug: these must differ here"
-        );
-    }
-
-    #[test]
-    fn a_page_folds_into_one_settled_assistant_message() {
-        let values = vec![
-            line(1, serde_json::json!({"type": "userMessage", "text": "hi"})),
-            line(
-                2,
-                serde_json::json!({"type": "assistantText", "delta": "hel"}),
-            ),
-            line(
-                3,
-                serde_json::json!({"type": "assistantText", "delta": "lo"}),
-            ),
-        ];
-        let items = fold_page(&values).unwrap();
-        assert_eq!(
-            items,
-            vec![
-                TranscriptItem::UserMsg {
-                    seq: 1,
-                    text: "hi".to_string(),
-                    attachments: Vec::new(),
-                },
-                TranscriptItem::AssistantMsg {
-                    seq: 2,
-                    text: "hello".to_string(),
-                    settled: false,
-                },
-            ]
-        );
-    }
-
-    #[test]
-    fn an_unparseable_line_fails_the_whole_page() {
-        let values = vec![serde_json::json!({"seq": 1, "ts": 1.0, "type": "not-a-real-type"})];
-        let err = fold_page(&values).unwrap_err();
-        assert!(err.contains("couldn't parse"));
-    }
-}
@@ -6,6 +6,7 @@ use android_view::{
 };
 use iris_core::{UiData, UiRenderNode, UiRenderState};
 use pollster::FutureExt;
+use std::time::{Duration, Instant};
 use wgpu::{
    rwh::{DisplayHandle, HandleError, HasDisplayHandle, HasWindowHandle, WindowHandle},
    *,
@@ -51,8 +52,19 @@ pub struct AndroidRenderer {

 impl AndroidRenderer {
    pub fn new(window: NativeWindow, width: u32, height: u32) -> Self {
+        // `force-gles` (RUST.md's I5 "Where iris's frame time goes") swaps
+        // the software-Vulkan (SwiftShader) path for GLES/virgl on the same
+        // build, to isolate whether the backend itself explains the frame
+        // time gap against Compose. `cfg!` rather than a runtime switch:
+        // there is no way to hand an env var to an already-launched Android
+        // process on this machine (see the feature's doc in Cargo.toml).
+        let backends = if cfg!(feature = "force-gles") {
+            Backends::GL
+        } else {
+            Backends::PRIMARY
+        };
        let instance = Instance::new(&InstanceDescriptor {
-            backends: Backends::PRIMARY,
+            backends,
            ..Default::default()
        });

@@ -74,12 +86,11 @@ impl AndroidRenderer {

        // Same request as the winit backend's `UiRenderer::new` -- no
        // binding-array features, see TEXTURES.md's "Recommended shape".
+        // `iris_core::device_limits()` is shared between the two backends;
+        // see its own doc for why it is not simply `Limits::default()`.
        let (device, queue) = adapter
            .request_device(&DeviceDescriptor {
-                required_limits: Limits {
-                    max_buffer_size: 1 << 30,
-                    ..Default::default()
-                },
+                required_limits: iris_core::device_limits(),
                ..Default::default()
            })
            .block_on()
@@ -128,7 +139,18 @@ impl AndroidRenderer {
        self.ui.update(&self.device, &self.queue, ui, render);
    }

-    pub fn draw(&mut self) {
+    /// Draws and presents one frame, returning the time spent in
+    /// `queue.submit` plus `present()` -- wherever a driver/GPU/compositor
+    /// wait would actually show up. The caller (`android::view::render`)
+    /// already times the whole frame from its own `redraw_to_submit` start;
+    /// subtracting this from that total is `redraw_to_submit` itself
+    /// (layout, text, primitive building, and this method's own render-pass
+    /// recording). RUST.md's I5 "Where iris's frame time goes" diagnosis,
+    /// added 2026-09-05 -- see `iris_core::FrameReport::record_split`'s own
+    /// doc for the caveat this shares: `present()` is not fenced against
+    /// the GPU actually finishing, so this is "how long the CPU was blocked
+    /// handing the frame off", not confirmed GPU time.
+    pub fn draw(&mut self) -> Duration {
        let output = self.surface.get_current_texture().unwrap();
        let view = output
            .texture
@@ -151,8 +173,10 @@ impl AndroidRenderer {
            self.ui.draw(render_pass);
        }

+        let submit_start = Instant::now();
        self.queue.submit(std::iter::once(encoder.finish()));
        output.present();
+        submit_start.elapsed()
    }

    pub fn size(&self) -> iris_core::util::Vec2 {
@@ -169,9 +193,22 @@ impl AndroidRenderer {

 /// `Tasks`' redraw handle on Android: a background task finishes on the
 /// tokio thread `Tasks::init` spawned, which is not attached to the JVM, so
-/// asking for a frame means attaching first. `post_frame_callback` needs a
-/// live `View` reference; the global ref is what survives past the JNI call
-/// that handed it to us.
+/// asking for a frame means attaching first. The global ref is what
+/// survives past the JNI call that handed the `View` to us.
+///
+/// **Goes through `View::post_delayed`, not `post_frame_callback`
+/// directly** -- found the hard way (RUST.md's I5 Android integration):
+/// `post_frame_callback`'s Java side calls `Choreographer.getInstance()`,
+/// which throws `IllegalStateException` unless the *calling* thread already
+/// has a `Looper` (`Choreographer.getInstance()`'s own contract). A tokio
+/// worker thread, even freshly attached to the JVM, has none -- the crash
+/// was a `JavaException` inside `View::post_frame_callback`'s `.unwrap()`,
+/// aborting the process on the second `redraw.request_redraw()` any
+/// android transcript-screen fetch made. `View.postDelayed(Runnable, 0)`
+/// is the ordinary Android answer to "queue work onto a View's own UI
+/// thread from any thread" and needs no Looper of its own; `delayed_callback`
+/// below is what that Runnable resolves to on the UI thread, where a real
+/// `post_frame_callback` is safe again.
 pub struct AndroidRedrawHandle {
    vm: JavaVM,
    view: GlobalRef,
@@ -189,6 +226,6 @@ impl RequestRedraw for AndroidRedrawHandle {
            return;
        };
        let local = env.new_local_ref(&self.view).unwrap();
-        View(local).post_frame_callback(&mut env);
+        View(local).post_delayed(&mut env, 0);
    }
 }
@@ -4,7 +4,7 @@ use accesskit_android::Adapter as AccessAdapter;
 use android_view::{
    AccessibilityNodeInfo, AccessibilityNodeProvider, Bundle, CallbackCtx, Context,
    InputConnection, KeyEvent, MotionEvent, Rect, View, ViewPeer,
-    jni::{JNIEnv, sys::jint},
+    jni::{JNIEnv, JavaVM, objects::GlobalRef, sys::jint},
    ndk::event::{Keycode, MotionAction},
 };
 // `marker::Sized` explicitly: `crate::prelude::*` below also brings in the
@@ -57,6 +57,11 @@ pub struct AndroidUiState {
    /// The AccessKit tree itself -- see `iris_core::AccessTree`'s doc
    /// comment.
    pub access: AccessTree,
+    /// iris's own frame-time report (RUST.md's I5 box, "Measurements
+    /// taken" (b)) -- `render()` below records into it once per frame,
+    /// because `dumpsys gfxinfo` cannot see a `SurfaceView`'s own
+    /// GPU-drawn frames at all. See `iris_core::FrameReport`'s own doc.
+    pub frame_report: FrameReport,
 }

 impl AndroidUiState {
@@ -72,6 +77,7 @@ impl AndroidUiState {
            shared,
            access_adapter: Default::default(),
            access: AccessTree::new(),
+            frame_report: FrameReport::new(),
        }
    }

@@ -101,6 +107,19 @@ pub trait AndroidAppState: HasAndroidUiState {
    fn back_pressed(&mut self, rsc: &mut AndroidRsc<Self>, render: &mut UiRenderState) -> bool {
        false
    }
+    /// Called once, right after `new`, with a fresh `JavaVM` handle and a
+    /// global reference to this app's own `View` -- for a caller that
+    /// needs to call into Java itself beyond what a [`RequestRedraw`]
+    /// handle already covers (P0's bench build calling
+    /// `BatteryManager`/`ClipboardManager` through the view's `Context`,
+    /// docs/RUST.md). Not folded into `new` itself: most implementors need
+    /// nothing here, and `new`'s job is building the widget tree, not
+    /// holding a platform handle -- the default does nothing. `vm`/`view`
+    /// are independent handles from the ones `new_peer` keeps for its own
+    /// `RequestRedraw` (a fresh `get_java_vm`/`new_global_ref` each), so
+    /// storing them has no effect on that mechanism.
+    #[allow(unused_variables)]
+    fn platform_ready(&mut self, rsc: &mut AndroidRsc<Self>, vm: JavaVM, view: GlobalRef) {}
 }

 /// The android-view analogue of `default::DefaultRsc` -- identical in
@@ -262,6 +281,13 @@ impl<State: AndroidAppState> IrisViewPeer<State> {
                .and_then(|r| self.render.window_region(r, &self.rsc)),
            self.window_size(),
        );
+        // iris's own frame-time report (RUST.md's I5 box, "Measurements
+        // taken" (b)): started here, at the same point a redraw request
+        // fires, and stopped after `renderer.draw()`'s `queue.submit` +
+        // `present()` -- the span Compose's render report and `gfxinfo`
+        // both count. See `iris_core::FrameReport`'s own doc for exactly
+        // what this does and does not measure.
+        let frame_start = Instant::now();
        let ui_state = self.state.android_state_mut();
        self.render.update(&ui_state.root, &mut self.rsc);
        let ui_state = self.state.android_state_mut();
@@ -269,7 +295,11 @@ impl<State: AndroidAppState> IrisViewPeer<State> {
            return;
        };
        renderer.update(&mut self.rsc.ui, &mut self.render);
-        renderer.draw();
+        let submit_to_present = renderer.draw();
+        self.state
+            .android_state_mut()
+            .frame_report
+            .record_split(frame_start.elapsed(), submit_to_present);
        let ui_state = self.state.android_state();
        log::debug!(
            "render(): after update active={} root_px={:?}",
@@ -432,6 +462,20 @@ impl<State: AndroidAppState> ViewPeer for IrisViewPeer<State> {
        self.render(ctx);
    }

+    /// Where `AndroidRedrawHandle::request_redraw` (`android/render.rs`)
+    /// actually lands: `View.postDelayed`'s Runnable resolves to this, on
+    /// the UI thread, which is what makes it safe to call from a background
+    /// task's own thread when `post_frame_callback`'s `Choreographer`
+    /// requirement (a `Looper` on the *calling* thread) is not. Same body
+    /// as `do_frame` -- draining tasks and rendering immediately is a
+    /// perfectly good answer to "a background fetch has new state," and
+    /// avoids a second frame-scheduling path to keep in sync with the real
+    /// one.
+    fn delayed_callback(&mut self, ctx: &mut CallbackCtx) {
+        self.drain_tasks();
+        self.render(ctx);
+    }
+
    fn as_input_connection(&mut self) -> Option<&mut dyn InputConnection> {
        Some(self)
    }
@@ -530,7 +574,10 @@ pub fn new_peer<'local, State: AndroidAppState>(
    };
    let shared = Rc::new(RefCell::new(Shared::default()));
    let ui_state = AndroidUiState::new(shared.clone());
-    let state = State::new(ui_state, &mut rsc);
+    let mut state = State::new(ui_state, &mut rsc);
+    let platform_vm = env.get_java_vm().unwrap();
+    let platform_view = env.new_global_ref(&view.0).unwrap();
+    state.platform_ready(&mut rsc, platform_vm, platform_view);
    let peer = IrisViewPeer {
        rsc,
        render: UiRenderState::new(),
@@ -102,13 +102,12 @@ impl UiRenderer {
        // needs descriptor indexing. See TEXTURES.md's "Recommended shape"
        // for why the old binding array asked for
        // VK_EXT_descriptor_indexing unconditionally and did not survive a
-        // real share of Android GPUs.
+        // real share of Android GPUs. `iris_core::device_limits()` is
+        // shared with the Android backend; see its own doc for why it is
+        // not simply `Limits::default()`.
        let (device, queue) = adapter
            .request_device(&DeviceDescriptor {
-                required_limits: Limits {
-                    max_buffer_size: 1 << 30,
-                    ..Default::default()
-                },
+                required_limits: iris_core::device_limits(),
                ..Default::default()
            })
            .block_on()
@@ -444,8 +444,31 @@ impl DragArbiter {
        self.state = ArbiterState::Undecided { already_selected };
    }

+    /// Whether this arbiter has no press in flight -- either it has never
+    /// seen [`Self::press_start`], or the last one it saw was released.
+    /// What a caller whose own `PressStart` sense can miss (see
+    /// [`Self::update`]'s doc) uses to notice that a `Pressing` frame has
+    /// arrived with no matching start, and recover by starting one now.
+    pub fn is_idle(&self) -> bool {
+        matches!(self.state, ArbiterState::Idle)
+    }
+
    /// The press continues (still down) at `pos`. Call once per frame
    /// while the button/finger is down; returns what this frame means.
+    ///
+    /// A caller must not call this while [`Self::is_idle`] is true for a
+    /// press that is genuinely still down -- `Idle` has no way to tell
+    /// "no press is happening" from "a press is happening but this
+    /// arbiter never got its `press_start`," so it always answers
+    /// `Undecided` and never leaves `Idle` on its own. That second case is
+    /// real: a touch's `ACTION_DOWN` lands whatever pixel the finger
+    /// actually hit, which is not guaranteed to be inside the same
+    /// row-local sensor region a later `ACTION_MOVE` in the same gesture
+    /// lands in (a row's own padding/gap, or its non-selectable header, is
+    /// pointer-transparent to `CursorSense`) -- so the widget that
+    /// receives the gesture's first `Pressing` frame may never have seen
+    /// its `PressStart`. `iris::transcript_ui::selection::Selection::drag`
+    /// is the caller that recovers from this, via `is_idle`.
    pub fn update(&mut self, pos: Vec2, now: Instant) -> DragOutcome {
        match self.state {
            ArbiterState::Idle => DragOutcome::Undecided,
@@ -584,6 +607,50 @@ mod drag_arbiter_tests {
        );
    }

+    /// A fresh arbiter that never saw `press_start` -- the state a widget's
+    /// own arbiter is left in when a gesture's `ACTION_DOWN` landed on a
+    /// pixel no sensor covered (a row's padding/gap, or its header) and
+    /// only a later `ACTION_MOVE` reached this widget. `update` must not
+    /// silently swallow the whole rest of the gesture here; `is_idle` is
+    /// what a caller checks to notice and recover (RUST.md's I5
+    /// intermittent-touch-scroll-dropout finding, 2026-09-05) --
+    /// `transcript_ui::selection::Selection::drag` is the real caller,
+    /// this is the pure-state half of the fix.
+    #[test]
+    fn is_idle_reports_a_press_that_was_never_started() {
+        let a = DragArbiter::new();
+        assert!(a.is_idle());
+    }
+
+    #[test]
+    fn update_on_an_idle_arbiter_stays_undecided_forever_without_recovery() {
+        // Documents the failure this fix works around: calling `update`
+        // (as if the arbiter were mid-gesture) without ever having called
+        // `press_start` leaves it stuck answering `Undecided`, even for a
+        // movement well past `DRAG_SLOP` that would otherwise pan
+        // immediately.
+        let mut a = DragArbiter::new();
+        assert_eq!(
+            a.update(Vec2::new(0.0, 100.0), t(10)),
+            DragOutcome::Undecided
+        );
+        assert!(a.is_idle());
+    }
+
+    #[test]
+    fn a_caller_can_recover_a_missed_press_start_via_is_idle() {
+        let mut a = DragArbiter::new();
+        // Simulates the real call site: a `Pressing` frame arrives with no
+        // matching `PressStart` ever having reached this arbiter.
+        assert!(a.is_idle());
+        a.press_start(Vec2::new(0.0, 700.0), t(0), false);
+        assert_eq!(
+            a.update(Vec2::new(0.0, 720.0), t(10)),
+            DragOutcome::Pan(20.0)
+        );
+        assert!(!a.is_idle());
+    }
+
    #[test]
    fn release_resets_to_idle() {
        let mut a = DragArbiter::new();
@@ -72,6 +72,18 @@ impl<Rsc: HasState> Tasks<Rsc> {
        )
    }

+    /// The same redraw handle `spawn`'s wrapper calls once, after a whole
+    /// task's future completes -- exposed so a caller running its own
+    /// longer-lived loop *inside* a spawned task (a live SSE follow, here)
+    /// can ask for a frame after each `TaskCtx::update`, not just at the
+    /// end. Without this a caller has no way to get a redraw mid-stream,
+    /// which is exactly the gap `iris/desktop-app`'s `app.rs` module doc
+    /// names for why it uses winit's `Proxy` instead of `Tasks` -- Android
+    /// has no `Proxy`, so this is what closes the same gap there.
+    pub fn redraw_handle(&self) -> Arc<dyn RequestRedraw> {
+        self.redraw.clone()
+    }
+
    pub fn spawn<F: AsyncFnOnce(TaskCtx<Rsc>) + 'static + std::marker::Send>(&mut self, task: F)
    where
        F::CallOnceFuture: Send,
@@ -18,3 +18,9 @@ iris = { path = ".." }
 client-core = { path = "../../client-core" }
 event-model = { path = "../../event-model" }
 pulldown-cmark = { workspace = true }
+# Selection has no accessibility label of its own yet (IRIS_TODO.md's
+# "Row-level accessibility names"), so a logcat line at
+# begin/extend is the smallest way to confirm a real long-press-then-drag
+# reached `Selection` on-device (RUST.md's I5 box, "Measurements taken"
+# (c)) -- version pinned to match `iris-android-app`'s own dependency.
+log = "0.4.28"
@@ -184,13 +184,40 @@ impl Selection {
                self.arbiter.release();
                return;
            }
+            // A `Pressing` frame with the arbiter still `Idle` means this
+            // gesture's `ACTION_DOWN` landed somewhere no row's sensor
+            // covers (a row's own padding/gap, or a header with no
+            // selection handler) and this row is only now getting the
+            // touch as it moves across it -- the touch is definitely still
+            // down (that's what `Pressing` means), so without this the
+            // arbiter would sit in `Idle` answering `Undecided` for the
+            // rest of the gesture (`DragArbiter::update`'s own doc).
+            // Recovered by starting the press here instead of where it
+            // was missed -- RUST.md's I5 intermittent-touch-scroll-dropout
+            // finding, 2026-09-05.
+            _ if self.arbiter.is_idle() => {
+                let already_selected = self.has_selection(ui);
+                self.arbiter.press_start(pos_window, now, already_selected);
+                self.arbiter.update(pos_window, now)
+            }
            _ => self.arbiter.update(pos_window, now),
        };
        match outcome {
            DragOutcome::Undecided => {}
            DragOutcome::Pan(dy) => list(ui).scroll(-dy),
-            DragOutcome::SelectStart => self.begin(ui, key, pos_row, size),
-            DragOutcome::SelectExtend => self.extend(ui, key, pos_row, size),
+            DragOutcome::SelectStart => {
+                // Grep-able on "iris selection" the way the frame report is
+                // on "iris frame report" -- selection has no accessibility
+                // label of its own yet, so this is the smallest way to
+                // confirm a real on-device long-press-then-drag actually
+                // reached here (RUST.md's I5 box, "Measurements taken" (c)).
+                log::info!("iris selection: begin at row {key:?}");
+                self.begin(ui, key, pos_row, size);
+            }
+            DragOutcome::SelectExtend => {
+                log::info!("iris selection: extend to row {key:?}");
+                self.extend(ui, key, pos_row, size);
+            }
        }
    }

@@ -262,6 +289,59 @@ mod tests {
        }
    }

+    /// The RUST.md I5 intermittent-touch-scroll-dropout regression: a
+    /// gesture whose `ACTION_DOWN` landed where no row's sensor covers
+    /// (padding, a gap, a header with no handler) delivers this row only
+    /// `Pressing` frames, never `PressStart`. Before the fix, the shared
+    /// `DragArbiter` stayed `Idle` for the whole gesture (`DragArbiter::
+    /// update`'s own doc), which is exactly what a real device trace
+    /// showed for four of twenty-four otherwise-identical swipes in one
+    /// run -- `ui-trace`-driven touch coordinates land on different row
+    /// content each time the list actually scrolls, so whether `DOWN`
+    /// happens to hit a sensor is intermittent by construction. This test
+    /// fails on the code before `Selection::drag`'s `_ if self.arbiter.
+    /// is_idle()` branch existed, because the arbiter would still report
+    /// `is_idle()` after both calls below.
+    #[test]
+    fn a_missed_press_start_recovers_on_the_next_pressing_frame() {
+        let mut rsc = TestRsc {
+            ui: UiData::default(),
+        };
+        let field = rsc
+            .ui
+            .widgets
+            .add_strong(TextEdit::new(
+                TextView::new(TextBuffer::new_empty(), TextAttrs::default(), None),
+                EditMode::MultiLine,
+            ))
+            .weak();
+        let list = rsc.ui.widgets.add_strong(List::new(Axis::Y)).weak();
+
+        let mut sel = Selection::new();
+        sel.register(1, field);
+        assert!(sel.arbiter.is_idle());
+
+        let now = Instant::now();
+        let size = Vec2::new(100.0, 20.0);
+        // No `PressStart` is ever sent -- only the `Pressing` frames a
+        // widget whose sensor missed the `ACTION_DOWN` would actually see.
+        sel.drag(
+            &mut rsc,
+            list,
+            1,
+            Vec2::ZERO,
+            size,
+            Vec2::new(540.0, 700.0),
+            CursorSense::Pressing(CursorButton::Left),
+            now,
+        );
+        assert!(
+            !sel.arbiter.is_idle(),
+            "a Pressing frame with the arbiter still Idle must recover \
+             the press rather than leaving it stuck"
+        );
+    }
+
    #[test]
    fn unregister_forgets_the_row_and_clears_a_matching_anchor() {
        let mut rsc = TestRsc {
@@ -35,6 +35,34 @@ fn iris_features() -> Features {
 /// kept for the big storage buffers behind rects/glyphs.
 const IRIS_MAX_BUFFER_SIZE: u64 = 1 << 30;

+/// Mirrors `iris_core::device_limits()` (`iris/core/src/render/mod.rs`) --
+/// cannot call it directly, since this rig is deliberately its own crate,
+/// not a workspace member (this file's own Cargo.toml comment). Keep the
+/// two in sync by hand when one changes; this rig's whole purpose is "does
+/// the device iris actually builds come back," so a stale copy here would
+/// silently stop answering that question. Zeroed rather than left at
+/// `Limits::default()`'s desktop-tier values because nothing in iris
+/// creates a `ComputePipeline` or a `@compute` shader stage -- found by
+/// grepping the whole `iris`/`iris-core` tree before this rig's comment was
+/// written -- and the unconditional default request is what crashed
+/// `request_device` on the Android emulator's software GL path
+/// (`EMU_GPU=software`, `force-gles`: SwiftShader's GL reports itself as
+/// OpenGL ES 3.0, which has no compute shaders at all, so the adapter's
+/// real limit is 0). The same would happen on a real GLES-3.0-only Android
+/// device.
+fn iris_limits() -> Limits {
+    Limits {
+        max_buffer_size: IRIS_MAX_BUFFER_SIZE,
+        max_compute_workgroup_storage_size: 0,
+        max_compute_invocations_per_workgroup: 0,
+        max_compute_workgroup_size_x: 0,
+        max_compute_workgroup_size_y: 0,
+        max_compute_workgroup_size_z: 0,
+        max_compute_workgroups_per_dimension: 0,
+        ..Default::default()
+    }
+}
+
 fn main() {
    vk::report();

@@ -104,10 +132,7 @@ fn main() {
    // limits requested, this is expected to succeed everywhere -- this rig
    // is what turned that from an assumption into a measurement, first on
    // this emulator's software Vulkan.
-    let wanted = Limits {
-        max_buffer_size: IRIS_MAX_BUFFER_SIZE,
-        ..Default::default()
-    };
+    let wanted = iris_limits();
    match pollster::block_on(adapter.request_device(&DeviceDescriptor {
        required_features: iris_features(),
        required_limits: wanted,
@@ -36,6 +36,7 @@ dependencies = [
 "sha2",
 "tempfile",
 "thiserror",
+ "time",
 "tokio",
 "tokio-stream",
 "tower",
@@ -57,6 +57,12 @@ ureq = { version = "3", features = ["json"] }
 # both in the graph rustls refuses to auto-select one.
 rustls = "0.23"
 libc = "0.2.189"
+# One ISO-8601 timestamp: the reset time on the invented rate-limit window
+# an echo session's `/usage` puts up. Already in the tree behind the
+# certificate machinery, so this is a direct name for what is compiled
+# anyway rather than a new crate -- and the alternative was hand-rolling a
+# civil-from-days conversion to print one line.
+time = { version = "0.3", features = ["formatting"] }

 [dev-dependencies]
 tempfile = "3"
@@ -207,6 +207,20 @@ mod tests {
    /// like any other.
    #[tokio::test]
    async fn a_spooled_enrollment_is_adopted_on_first_use() {
+        // Under a subscriber, like every other exercise of this middleware.
+        // `tracing` caches a callsite's interest process-wide the first time it
+        // is reached, so the refusal at the end of this test -- reached with no
+        // subscriber on this thread -- could cache the rejection warning as
+        // never-enabled and make the tripwire above see an empty log. That
+        // failed about one full-suite run in ten, in the test that exists to
+        // notice a credential leak, which is the worst place for a flake.
+        let _guard = tracing::subscriber::set_default(
+            tracing_subscriber::fmt()
+                .with_max_level(tracing::Level::TRACE)
+                .with_writer(std::io::sink)
+                .finish(),
+        );
+
        let dir = tempfile::tempdir().expect("tempdir");
        let manager = manager_with_token(dir.path(), "first");
        let spooled = generate_token();
@@ -30,6 +30,18 @@ pub struct Config {
    pub tokens: Vec<TokenEntry>,
    pub setups: Vec<SetupConfig>,
    pub sessions: Vec<SessionConfig>,
+    /// What a new session's thinking level is when nothing chose one.
+    ///
+    /// Here rather than on a provider because providers are *discovered*: a
+    /// default written onto one would be erased by the next rediscovery, which
+    /// is the kind of setting that looks like it stuck until the day it did
+    /// not. Here rather than on the phone because a second device would then
+    /// spawn sessions the first one's owner did not expect.
+    ///
+    /// `None` is the CLI's own default, and stays reachable: this is a level
+    /// somebody chose, not a level this app picked for them.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub default_effort: Option<String>,
 }

 /// A machine, and the things it can run.
@@ -78,6 +90,16 @@ pub struct ProviderConfig {
    pub models: Vec<String>,
 }

+impl ProviderConfig {
+    /// The executable to run for this provider: its override, or its kind's
+    /// default.
+    pub fn program(&self) -> &str {
+        self.command
+            .as_deref()
+            .unwrap_or(self.kind.default_program())
+    }
+}
+
 /// How to reach a setup that isn't this machine, with the system `ssh` client
 /// -- so `~/.ssh/config`, agents and jump hosts all keep working, and there is
 /// one place to configure connections. A remote session is the identical
@@ -94,6 +116,19 @@ pub struct SshConfig {
    /// Extra `-o` settings, each written as `Key=value`.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub options: Vec<String>,
+    /// Where this machine keeps the GGUF models it can serve, absent for
+    /// the same default this backend uses (`~/.local/share/ai-app/models`
+    /// -- `$XDG_DATA_HOME` is not read on the far side, since it is this
+    /// machine's environment that would answer). A `~` prefix is the
+    /// remote home.
+    ///
+    /// Here rather than on the provider because it is a fact about the
+    /// machine, and because a machine reached over ssh is where the model
+    /// has to be: a llama.cpp session serves the file from the machine
+    /// that runs `llama-server`, and this backend's own downloads are on
+    /// whichever machine that is only when they are the same one.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub models_dir: Option<PathBuf>,
    /// Where a file attached from the phone is put on this machine so the
    /// session can read it. Absent means the session's own working directory,
    /// or the login home for a session that has none. A `~` prefix is the
@@ -145,6 +180,45 @@ impl DriverKind {
        }
    }

+    /// Which paid service meters a session of this kind, and `None` for one
+    /// that costs nothing.
+    ///
+    /// What decides which account -- if any -- a rate-limit bar is about is the
+    /// provider a session runs, not the machine it runs on: an echo session on
+    /// a machine that also has the Claude CLI was drawn with that CLI's
+    /// five-hour window, a quota it cannot spend.
+    ///
+    /// Echo names a meter of its own that exists only when a test has asked for
+    /// one (`/usage` in `session::echo`), which is how the bar's states are
+    /// reached without an account. With none set there is no snapshot, and the
+    /// phone draws nothing.
+    ///
+    /// The string is a [`crate::usage::UsageProvider::name`], and it is what
+    /// pairs a session with one of `GET /usage`'s snapshots -- so
+    /// `usage::providers_for` reads this rather than matching on kinds again.
+    pub fn usage_provider(self) -> Option<&'static str> {
+        match self {
+            Self::ClaudeCli => Some(crate::usage::CLAUDE),
+            Self::Echo => Some(crate::usage::ECHO),
+            Self::LlamaCpp => None,
+        }
+    }
+
+    /// The executable a provider of this kind runs when it names none.
+    ///
+    /// Here rather than at each spawn site because it is not only the spawn
+    /// that runs it: `usage` runs the Claude CLI too, to have it refresh its
+    /// own OAuth token, and a default that disagreed with the driver's would
+    /// ask the wrong binary on a machine with two installs.
+    pub fn default_program(self) -> &'static str {
+        match self {
+            Self::ClaudeCli => "claude",
+            Self::LlamaCpp => "llama-server",
+            // Echo is translated in-process; nothing is spawned for it.
+            Self::Echo => "echo",
+        }
+    }
+
    /// Whether the conversation exists outside this app, so that deleting the
    /// session here does not end it.
    ///
@@ -162,6 +236,24 @@ impl DriverKind {
            Self::Echo | Self::LlamaCpp => false,
        }
    }
+
+    /// Whether a thinking level means anything to this kind, so the phone can
+    /// offer the control only where it does something.
+    ///
+    /// Reported from here rather than decided on the phone, and asked of the
+    /// *kind* rather than branched on: the alternative is the session-type
+    /// `if` this app does not have anywhere else. `--effort` is the Claude
+    /// CLI's; a llama session's sampling is `params`, and echo does not think.
+    ///
+    /// It matters more than a control that would simply do nothing, because
+    /// choosing a level stops the process -- so on a session that cannot use
+    /// one it is a button whose only effect is the cost.
+    pub fn takes_effort(self) -> bool {
+        match self {
+            Self::ClaudeCli => true,
+            Self::Echo | Self::LlamaCpp => false,
+        }
+    }
 }

 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -196,6 +288,17 @@ pub struct SessionConfig {
    /// the CLI stays the one authority on which modes exist.
    #[serde(skip_serializing_if = "Option::is_none")]
    pub permission_mode: Option<String>,
+    /// How hard the model thinks, passed straight to `--effort`. A string for
+    /// the same reason `permission_mode` is: the CLI owns which levels exist.
+    ///
+    /// Unlike the model and the mode, there is no control request that changes
+    /// one -- checked against 2.1.258, whose only two are `set_model` and
+    /// `set_permission_mode` -- so this is settled at launch and `None` means
+    /// whatever the CLI's own default is. That is a state the phone has to be
+    /// able to *choose*, not just start in, which is why it is an option
+    /// rather than a level with a default written here.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub effort: Option<String>,
    /// Settings the driver interprets, chosen at spawn.
    ///
    /// Deliberately untyped: what a temperature or a context size means is the
@@ -215,6 +318,30 @@ pub struct SessionConfig {
    /// turned off in one tap where one that never arrived is not diagnosable.
    #[serde(default = "notify_default")]
    pub notify: bool,
+    /// Whether a session stopped by the account's usage limit sends itself a
+    /// message once the limit lifts, instead of waiting for a person.
+    ///
+    /// Off unless somebody asked for it. It spends quota the moment it becomes
+    /// available and it does so while nobody is looking, which is exactly the
+    /// kind of thing that must not happen because a default said so.
+    #[serde(default, skip_serializing_if = "not_set")]
+    pub auto_resume: bool,
+    /// What that message says. `None` is [`DEFAULT_RESUME_MESSAGE`], and stays
+    /// reachable: it is this app's word, not one somebody chose, so clearing
+    /// the field goes back to it rather than sending an empty message.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub auto_resume_message: Option<String>,
+    /// The message this session owes itself once the limit lifts, and when to
+    /// try. Written when a limit is hit, moved when the wait turns out to be
+    /// wrong, and cleared when the message goes out or auto-resume is turned
+    /// off -- see [`ScheduledResume`].
+    ///
+    /// Persisted rather than held in memory because the wait outlives the
+    /// process doing it: a five-hour window and a weekly one both routinely
+    /// outlast a backend restart, and a resume forgotten across one is a
+    /// session that silently never comes back.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub resume: Option<ScheduledResume>,
    /// Whether this session's process is stopped when the server exits, instead
    /// of being left running for the next start to adopt.
    ///
@@ -231,6 +358,28 @@ pub struct SessionConfig {
    pub created: f64,
 }

+/// A message owed to a session whose account ran out, and when to try sending
+/// it.
+///
+/// `since` is the whole reason this is a struct: the wait is rescheduled every
+/// time the meter is asked and still says no, so `at` alone cannot say how long
+/// this has been going on -- and something has to, or a machine that can never
+/// be asked is retried until somebody notices. See `crate::resume`.
+#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct ScheduledResume {
+    /// Epoch seconds: when the limit is next worth checking. Never a promise
+    /// that the message goes out then -- the meter is asked first.
+    pub at: f64,
+    /// Epoch seconds the limit was hit.
+    pub since: f64,
+}
+
+/// What an auto-resume says when nothing else was chosen. One word, because
+/// the session already knows what it was doing and this is only the nudge that
+/// lets it carry on.
+pub const DEFAULT_RESUME_MESSAGE: &str = "continue";
+
 fn notify_default() -> bool {
    true
 }
@@ -385,6 +534,7 @@ mod tests {
                        port: Some(2222),
                        identity_file: None,
                        options: Vec::new(),
+                        models_dir: None,
                        attachments_dir: None,
                    }),
                    providers: vec![ProviderConfig {
@@ -395,6 +545,7 @@ mod tests {
                    }],
                },
            ],
+            default_effort: Some("low".to_string()),
            sessions: vec![SessionConfig {
                id: "abc123".to_string(),
                setup: "vm".to_string(),
@@ -403,8 +554,12 @@ mod tests {
                model: None,
                cwd: None,
                permission_mode: None,
+                effort: None,
                params: BTreeMap::new(),
                notify: true,
+                auto_resume: false,
+                auto_resume_message: None,
+                resume: None,
                throwaway: false,
                created: 1234.5,
            }],
@@ -16,6 +16,7 @@ mod config;
 mod files;
 mod media;
 mod models;
+mod resume;
 mod routes;
 mod session;
 mod setups;
@@ -264,7 +265,16 @@ async fn main() -> Result<()> {
    // No providers listed here any more: which machines can be asked, and about
    // what, comes from the setups at the moment the screen is opened -- so a
    // machine added from the phone reports its limits without a restart.
-    let monitor = Arc::new(usage::UsageMonitor::new());
+    // The fixture is the manager's, because that is where the `/usage` command
+    // that sets it is typed; the monitor is what serves it.
+    let monitor = Arc::new(usage::UsageMonitor::new(manager.usage_fixture()));
+
+    // The one thing in here that acts without a request behind it: a session
+    // switched to auto-resume waits out its account's usage limit and picks
+    // itself back up. Started whether or not any session has it on, because
+    // the setting is per session and changes from the phone -- see
+    // `resume::run`.
+    tokio::spawn(resume::run(Arc::clone(&manager), Arc::clone(&monitor)));

    // The bearer-token middleware wraps the entire router -- routes and fallback
    // alike -- here and only here, so a new route can't forget auth.
@@ -29,6 +29,8 @@ use serde::Serialize;

 use wg_app_link::private;

+use crate::session::transport::{Launch, Transport};
+
 /// Identifies this client to HuggingFace. They ask for one, and a request
 /// without it is more likely to be rate-limited.
 const USER_AGENT: &str = concat!("ai-server/", env!("CARGO_PKG_VERSION"));
@@ -521,6 +523,83 @@ fn collect(root: &Path, dir: &Path, found: &mut Vec<LocalModel>) {
    }
 }

+/// Where a machine reached over ssh keeps its models, when its setup does
+/// not say.
+///
+/// The same place this backend puts its own downloads, written out rather
+/// than derived: `$XDG_DATA_HOME` here describes *this* machine's
+/// environment, and the far machine's is the far machine's business. A
+/// setup whose models are elsewhere says so (`SshConfig::models_dir`).
+const FAR_MODELS_DIR: &str = "~/.local/share/ai-app/models";
+
+/// Which directory holds the models on the machine `transport` reaches.
+///
+/// One answer, because two things ask: the list a spawn screen offers,
+/// and the path a session hands `llama-server`. A machine that listed one
+/// directory and served from another would offer models that then failed
+/// to load, which reads as the model being broken.
+pub fn dir_on(transport: &Transport, local: &Path) -> String {
+    match transport {
+        Transport::Here => local.to_string_lossy().into_owned(),
+        Transport::Ssh { ssh, .. } => ssh
+            .models_dir
+            .as_ref()
+            .map_or(FAR_MODELS_DIR.to_string(), |dir| {
+                dir.to_string_lossy().into_owned()
+            }),
+    }
+}
+
+/// Every GGUF on the machine a setup names, which is the machine that
+/// would have to serve it.
+///
+/// The local half of this is [`ModelStore::list`], reading the same shape
+/// off this machine's disk; a caller picks by transport, since a setup
+/// with no ssh *is* this machine and asking a shell about it would be a
+/// slower way to the same answer. What must not happen is offering this
+/// backend's downloads for a session on another machine: the file has to
+/// be where `llama-server` runs, and a list that says otherwise is a
+/// claim about the wrong filesystem.
+///
+/// `dir` is that machine's models directory, `~` included -- expanded on
+/// the far side, which is the only place that knows what it is. A
+/// directory that is not there is an empty list rather than a failure: a
+/// machine that has never had a model put on it is an ordinary state, and
+/// the same one as a machine whose directory exists and is empty.
+pub async fn on_machine(transport: &Transport, dir: &str) -> Result<Vec<LocalModel>> {
+    let script = "p=$1; case $p in \"~\") p=$HOME;; \"~/\"*) p=$HOME/${p#\"~/\"};; esac; \
+                  [ -d \"$p\" ] || exit 0; \
+                  find \"$p\" -type f -name '*.gguf' -printf '%s\\t%P\\0'";
+    let launch = Launch::new(
+        "sh",
+        vec![
+            "-c".to_string(),
+            script.to_string(),
+            "sh".to_string(),
+            dir.to_string(),
+        ],
+        None,
+    );
+    let out = transport.capture(&launch).await?;
+    let mut found: Vec<LocalModel> = out
+        .split('\0')
+        .filter(|record| !record.is_empty())
+        // Two fields, and the name last, so a `\t` in a filename survives.
+        .filter_map(|record| record.split_once('\t'))
+        .filter_map(|(bytes, key)| {
+            let (repo, file) = key.rsplit_once('/')?;
+            Some(LocalModel {
+                key: key.to_string(),
+                repo: repo.to_string(),
+                file: file.to_string(),
+                bytes: bytes.trim().parse().unwrap_or(0),
+            })
+        })
+        .collect();
+    found.sort_by(|a, b| a.key.cmp(&b.key));
+    Ok(found)
+}
+
 /// A model repository on HuggingFace, as the browse screen shows it.
 #[derive(Debug, Clone, Serialize)]
 #[serde(rename_all = "camelCase")]
@@ -0,0 +1,382 @@
+//! Auto-resume: picking a session back up when its account's usage limit
+//! lifts.
+//!
+//! Off unless a session was switched to it, because this spends quota the
+//! moment quota exists and does it while nobody is watching. What it does is
+//! narrow on purpose: it sends one message -- "continue" unless something else
+//! was typed -- to a session that stopped because the account ran out, and
+//! then it is done. There is no retry loop around the conversation itself.
+//!
+//! **The schedule is a plan to ask, never a plan to send.** A reset time is
+//! the one thing here that cannot be trusted: the dialect's is a hint written
+//! when the turn failed, the endpoint's moves when the window moves, and both
+//! are wrong across the case this exists for -- a limit that lifts later than
+//! it said. So the wait ends in a *question* to [`crate::usage`], and only an
+//! answer that says the limits no longer apply sends anything. Every other
+//! answer, including one that cannot be got at all, becomes a new wait.
+//!
+//! This is the top layer: it holds the session manager and the usage monitor
+//! and neither holds it. That is what lets the decision below be a pure
+//! function of a snapshot and a clock, which is the whole of what is worth
+//! testing here.
+
+use std::sync::Arc;
+use std::time::Duration;
+
+use crate::session::{LimitHit, OwedResume, SessionManager, now};
+use crate::usage::{UsageMonitor, UsageSnapshot, UsageState};
+
+/// How often to look at the schedule. Coarse deliberately: a wait measured in
+/// hours does not deserve a fine-grained clock, and the meter behind it is
+/// cached for three minutes anyway.
+const TICK: Duration = Duration::from_secs(60);
+
+/// How close to a scheduled check is close enough to ask the meter. Anything
+/// further out is left alone, so a session waiting five hours costs nothing
+/// until the last few minutes of it.
+const NEARLY: f64 = 300.0;
+
+/// How long to wait after an answer that decided nothing -- the machine could
+/// not be asked, or it says the limit is still on with no reset time.
+const BACKOFF: f64 = 300.0;
+
+/// The least time to wait before asking again, whatever a reset time says. A
+/// window that claims to reset in the past would otherwise be asked about on
+/// every tick.
+const AT_LEAST: f64 = 60.0;
+
+/// How long after the limit was hit to stop waiting.
+///
+/// Something has to bound it, or a machine that can never be asked -- an
+/// unplugged laptop, a setup somebody edited away -- is retried for ever with
+/// nothing on screen saying so. A day is past the longest window Claude
+/// reports, so reaching this means the wait was never going to end on its own.
+const GIVE_UP: f64 = 24.0 * 60.0 * 60.0;
+
+/// The percentage at which a window is spent. The API counts up to 100, so
+/// this is an equality in all but name; written as a threshold because a
+/// figure arriving slightly over is a full window, not a corrupt one.
+const SPENT: f64 = 100.0;
+
+/// What to do about one owed resume, having asked the meter.
+#[derive(Debug, Clone, Copy, PartialEq)]
+pub enum Step {
+    /// The limits no longer apply: send the message.
+    Send,
+    /// Ask again at this epoch second.
+    WaitUntil(f64),
+    /// This has been waiting longer than anything real would take.
+    GiveUp,
+}
+
+/// Runs the schedule until the server stops.
+///
+/// Two things wake it: the tick, and a session reporting that it has just run
+/// out. The second is not an optimisation -- a limit hit is what *creates* a
+/// schedule, and a tick that happened a moment before it would otherwise leave
+/// the session unrecorded until the next one.
+pub async fn run(manager: Arc<SessionManager>, monitor: Arc<UsageMonitor>) {
+    let mut limits = manager.subscribe_limits();
+    loop {
+        tokio::select! {
+            _ = tokio::time::sleep(TICK) => {}
+            hit = limits.recv() => match hit {
+                Ok(LimitHit { session_id, resets_at }) => note(&manager, &session_id, resets_at),
+                // Lagged: some reports were dropped, and a session that hit a
+                // limit while this was busy has no schedule. Nothing is lost
+                // for good -- the sweep below reads the config, and the
+                // session will report again the next time it is poked -- but
+                // it is worth saying, because until then that session waits
+                // for a person.
+                Err(tokio::sync::broadcast::error::RecvError::Lagged(missed)) => {
+                    tracing::warn!("auto-resume missed {missed} limit reports");
+                }
+                Err(tokio::sync::broadcast::error::RecvError::Closed) => return,
+            },
+        }
+        sweep(&manager, &monitor).await;
+    }
+}
+
+/// Records a limit against the session that hit it, if it is one that resumes.
+pub(crate) fn note(manager: &SessionManager, session_id: &str, resets_at: Option<f64>) {
+    match manager.note_limit(session_id, resets_at) {
+        Ok(true) => tracing::info!("session {session_id} hit its usage limit; auto-resume is on"),
+        Ok(false) => {}
+        Err(err) => tracing::error!("couldn't schedule a resume for {session_id}: {err:#}"),
+    }
+}
+
+/// One pass over everything owed a message.
+async fn sweep(manager: &SessionManager, monitor: &Arc<UsageMonitor>) {
+    let at = now();
+    for owed in manager.owed_resumes() {
+        if owed.scheduled.at - at > NEARLY {
+            continue;
+        }
+        // Asked per session rather than once for the whole sweep: the answer
+        // is cached per machine and per meter, so several sessions on one
+        // account share one fetch, and a machine nobody is waiting on is not
+        // dialled at all.
+        let snapshot = snapshot_for(Arc::clone(monitor), manager, &owed).await;
+        match decide(snapshot.as_ref(), &owed, now()) {
+            Step::Send => match manager.resume_now(&owed.session_id) {
+                Ok(message) => tracing::info!(
+                    "the limit on {} has lifted; sent \"{message}\" to {}",
+                    owed.setup,
+                    owed.session_id
+                ),
+                Err(err) => {
+                    tracing::error!("couldn't resume {}: {err:#}", owed.session_id)
+                }
+            },
+            Step::WaitUntil(next) => {
+                if let Err(err) = manager.reschedule_resume(&owed.session_id, next) {
+                    tracing::error!(
+                        "couldn't move {}'s resume to {next}: {err:#}",
+                        owed.session_id
+                    );
+                }
+            }
+            Step::GiveUp => {
+                // About the machine rather than in the state's own words: the
+                // detail is in the log, and what lands in the transcript has
+                // to read on a phone.
+                let why = match snapshot.as_ref().map(|snapshot| &snapshot.state) {
+                    Some(UsageState::Ok) => "the limit has not lifted in a day".to_string(),
+                    _ => format!("{} could not be asked for a day", owed.setup),
+                };
+                if let Err(err) = manager.abandon_resume(&owed.session_id, &why) {
+                    tracing::error!("couldn't clear {}'s resume: {err:#}", owed.session_id);
+                }
+            }
+        }
+    }
+}
+
+/// The numbers for the machine and the meter this session is billed against,
+/// and `None` when nothing reports on it.
+///
+/// Blocking work, so it goes to a blocking thread: the fetch behind it reads a
+/// credential file over ssh and then makes an HTTP call.
+async fn snapshot_for(
+    monitor: Arc<UsageMonitor>,
+    manager: &SessionManager,
+    owed: &OwedResume,
+) -> Option<UsageSnapshot> {
+    let setups: Vec<_> = manager
+        .setups()
+        .into_iter()
+        .filter(|setup| setup.id == owed.setup)
+        .collect();
+    if setups.is_empty() {
+        return None;
+    }
+    let provider = owed.provider;
+    tokio::task::spawn_blocking(move || {
+        monitor
+            .snapshots(&setups)
+            .into_iter()
+            .find(|snapshot| snapshot.provider == provider)
+    })
+    .await
+    .unwrap_or_default()
+}
+
+/// What one owed resume should do, given what the meter said and the time.
+///
+/// A pure function of the two, which is what makes the rule inspectable: every
+/// answer that is not "the limits no longer apply" is a longer wait, and the
+/// only thing that ends the waiting other than success is the clock.
+///
+/// The reset time comes from the *snapshot* rather than from the schedule, so
+/// a window that turns out to reset later than the dialect said pushes the
+/// check back, and one that resets sooner pulls it forward. That is the case
+/// the whole design is about: the first answer was a guess, this one is a
+/// measurement.
+pub fn decide(snapshot: Option<&UsageSnapshot>, owed: &OwedResume, at: f64) -> Step {
+    let step = match snapshot {
+        // The meter answered with numbers, which is the only answer that can
+        // send anything.
+        Some(snapshot) if snapshot.state == UsageState::Ok => {
+            let spent: Vec<&crate::usage::UsageWindow> = snapshot
+                .windows
+                .iter()
+                .filter(|window| window.percent >= SPENT)
+                .collect();
+            if spent.is_empty() {
+                Step::Send
+            } else {
+                // The earliest of the spent windows: it is the first moment
+                // the situation can have changed, and if the others are still
+                // full this comes straight back here.
+                match spent
+                    .iter()
+                    .filter_map(|window| epoch_of(window.resets_at.as_deref()))
+                    .min_by(f64::total_cmp)
+                {
+                    Some(resets) => Step::WaitUntil(resets),
+                    // Spent with no reset time anybody could read. Not a
+                    // reason to send: what is known is that the limit is on.
+                    None => Step::WaitUntil(at + BACKOFF),
+                }
+            }
+        }
+        // Logged out, unreachable, or the endpoint refused us -- and nothing
+        // at all, which is a session whose machine or provider has gone. None
+        // of them says the limit has lifted, and sending on any of them is
+        // exactly the "inferred value presented as a measured one" this is
+        // built to avoid.
+        _ => Step::WaitUntil(at + BACKOFF),
+    };
+    match step {
+        // Waiting past the point where a real window would have reset means
+        // whatever is wrong is not going to fix itself.
+        Step::WaitUntil(_) if at - owed.scheduled.since > GIVE_UP => Step::GiveUp,
+        Step::WaitUntil(next) => Step::WaitUntil(next.max(at + AT_LEAST)),
+        other => other,
+    }
+}
+
+/// An RFC-3339 timestamp as epoch seconds, and `None` for one that is absent
+/// or unreadable -- the same two answers the phone's countdown makes, kept
+/// apart from each other nowhere here because both mean "this cannot decide
+/// when to ask".
+fn epoch_of(resets_at: Option<&str>) -> Option<f64> {
+    let text = resets_at?;
+    time::OffsetDateTime::parse(text, &time::format_description::well_known::Rfc3339)
+        .ok()
+        .map(|at| at.unix_timestamp() as f64)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::config::ScheduledResume;
+    use crate::usage::UsageWindow;
+
+    fn owed(since: f64) -> OwedResume {
+        OwedResume {
+            session_id: "s1".to_string(),
+            setup: "local".to_string(),
+            provider: crate::usage::CLAUDE,
+            scheduled: ScheduledResume { at: since, since },
+        }
+    }
+
+    fn snapshot(state: UsageState, windows: Vec<UsageWindow>) -> UsageSnapshot {
+        UsageSnapshot {
+            provider: crate::usage::CLAUDE.to_string(),
+            setup: "local".to_string(),
+            setup_name: "this machine".to_string(),
+            state,
+            windows,
+            fetched_at: 0.0,
+        }
+    }
+
+    fn window(percent: f64, resets_at: Option<&str>) -> UsageWindow {
+        UsageWindow {
+            kind: "session".to_string(),
+            label: "5-hour window".to_string(),
+            percent,
+            resets_at: resets_at.map(str::to_string),
+            active: true,
+        }
+    }
+
+    #[test]
+    fn a_meter_with_room_in_it_is_the_only_thing_that_sends() {
+        let clear = snapshot(UsageState::Ok, vec![window(41.0, None)]);
+        assert_eq!(decide(Some(&clear), &owed(0.0), 100.0), Step::Send);
+    }
+
+    #[test]
+    fn a_window_still_spent_moves_the_check_to_its_own_reset_time() {
+        // The case the feature exists for: the wait was scheduled for one
+        // time, the limit is still on, and the endpoint now names another.
+        let at = 1_788_546_972.0;
+        let later = "2026-09-05T12:00:00+00:00";
+        let spent = snapshot(UsageState::Ok, vec![window(100.0, Some(later))]);
+        assert_eq!(
+            decide(Some(&spent), &owed(at - 60.0), at),
+            Step::WaitUntil(epoch_of(Some(later)).expect("parses"))
+        );
+    }
+
+    #[test]
+    fn a_reset_time_already_past_still_waits_a_little() {
+        let at = 1_788_546_972.0;
+        let spent = snapshot(
+            UsageState::Ok,
+            vec![window(100.0, Some("2020-01-01T00:00:00+00:00"))],
+        );
+        assert_eq!(
+            decide(Some(&spent), &owed(at - 60.0), at),
+            Step::WaitUntil(at + AT_LEAST)
+        );
+    }
+
+    #[test]
+    fn the_earliest_spent_window_is_the_one_worth_waiting_on() {
+        let at = 1_788_546_972.0;
+        let soon = "2026-09-05T12:00:00+00:00";
+        let far = "2026-09-09T12:00:00+00:00";
+        let mut weekly = window(100.0, Some(far));
+        weekly.kind = "weekly_all".to_string();
+        let spent = snapshot(UsageState::Ok, vec![window(100.0, Some(soon)), weekly]);
+        assert_eq!(
+            decide(Some(&spent), &owed(at - 60.0), at),
+            Step::WaitUntil(epoch_of(Some(soon)).expect("parses"))
+        );
+    }
+
+    #[test]
+    fn a_meter_that_could_not_be_asked_never_sends() {
+        let at = 1_788_546_972.0;
+        for state in [
+            UsageState::NotLoggedIn,
+            UsageState::Unreachable {
+                detail: "no route".to_string(),
+            },
+            UsageState::Failed {
+                detail: "429".to_string(),
+            },
+        ] {
+            let broken = snapshot(state.clone(), Vec::new());
+            assert_eq!(
+                decide(Some(&broken), &owed(at - 60.0), at),
+                Step::WaitUntil(at + BACKOFF),
+                "{state:?}"
+            );
+        }
+        // And no snapshot at all -- a machine or provider edited away under a
+        // session that was waiting on it.
+        assert_eq!(
+            decide(None, &owed(at - 60.0), at),
+            Step::WaitUntil(at + BACKOFF)
+        );
+    }
+
+    #[test]
+    fn waiting_longer_than_any_real_window_gives_up_rather_than_retrying_for_ever() {
+        let at = 1_788_546_972.0;
+        let broken = snapshot(
+            UsageState::Unreachable {
+                detail: "no route".to_string(),
+            },
+            Vec::new(),
+        );
+        assert_eq!(
+            decide(Some(&broken), &owed(at - GIVE_UP - 1.0), at),
+            Step::GiveUp
+        );
+        // A meter that answers is still allowed to send on the same tick: the
+        // ceiling bounds waiting, not resuming.
+        let clear = snapshot(UsageState::Ok, vec![window(3.0, None)]);
+        assert_eq!(
+            decide(Some(&clear), &owed(at - GIVE_UP - 1.0), at),
+            Step::Send
+        );
+    }
+}
@@ -7,6 +7,7 @@
 //! POST   /setups                       add {name, ssh?} -- providers are discovered
 //! POST   /setups/probe                 dry run {ssh?}: what would be found there
 //! GET    /setups/{id}                  one machine, for refetching after a change
+//! GET    /setups/{id}/models           GGUFs on that machine, for a llama session
 //! GET    /setups/{id}/dir?path=P       entries of directory P, and P resolved
 //! GET    /setups/{id}/file?path=P      content of file P, or why not
 //! PUT    /setups/{id}/file             {path, content, ifSha256} -> new size/mtime/sha256
@@ -28,6 +29,12 @@
 //! GET    /sessions/{id}/transcript    a page of history: ?before=N (newest when absent),
 //!                                     ?limit=N, ?coalesce=true to count rows not deltas,
 //!                                     ?after=N to floor it at what the caller already holds
+//! GET    /sessions/{id}/subagents     [{id, title, status, created, lastActivity}], oldest
+//!                                     first -- see SUBAGENTS.md
+//! GET    /sessions/{id}/subagents/{sub}/transcript  exactly the transcript route above,
+//!                                     against that subagent's own transcript
+//! GET    /sessions/{id}/subagents/{sub}/events?after=N  exactly the events route above,
+//!                                     against that subagent's own stream
 //! POST   /sessions/{id}/message        {text, attachmentIds?}
 //!                                     (starts the process first if it has exited)
 //! POST   /sessions/{id}/unqueue        {messageId} -- take back one not read yet
@@ -41,6 +48,8 @@
 //!                                     which starts again in the new one
 //! POST   /sessions/{id}/model          {model}
 //! POST   /sessions/{id}/permission-mode  {permissionMode}
+//! POST   /sessions/{id}/effort         {effort} -- null for the CLI's default;
+//!                                     settled at launch, so this stops the process
 //! POST   /sessions/{id}/command        {text}  -- /compact, /clear, /rename x, or the dialect's own
 //!                                     (starts the process first if it has exited)
 //! POST   /sessions/{id}/compact
@@ -49,8 +58,12 @@
 //! DELETE /sessions/{id}                kill process, delete transcript + files
 //!                                     (?deleteForeign=true removes the machine's own copy too)
 //! POST   /sessions/{id}/notify         {notify}  -- announce this one or not
+//! POST   /sessions/{id}/auto-resume    {autoResume, message?} -- carry on by itself
+//!                                     once the account's usage limit lifts
 //! GET    /notifications                SSE: every session's attention-wanting
 //!                                      moments, live only (see `notifications`)
+//! GET    /defaults                    {effort} -- what a new session starts at
+//! POST   /defaults                    {effort} -- null for the CLI's own default
 //! GET    /usage                        cached usage windows per provider
 //! GET    /models                       downloaded GGUFs, and what is being fetched
 //! GET    /models/search?q=Q            HuggingFace repositories matching Q
@@ -92,6 +105,7 @@ use tokio_stream::wrappers::{BroadcastStream, ReceiverStream};

 use crate::session::driver::{SessionCommand, Unqueued};
 use crate::session::pending::Operation;
+use crate::session::subagent::{Subagent, SubagentInfo};
 use crate::session::transcript::{CATCH_UP_LIMIT, CatchUp, SeqEvent, catch_up};
 use crate::session::{LiveSession, SessionInfo, SessionManager, SpawnSpec};

@@ -109,6 +123,8 @@ pub fn router(manager: Arc<SessionManager>) -> Router {
            "/setups/{id}",
            get(read_setup).put(update_setup).delete(delete_setup),
        )
+        // The models on the machine a setup names, for a llama session there.
+        .route("/setups/{id}/models", get(setup_models))
        // The filesystem of the machine a setup names. Under the setup
        // rather than under a session because a filesystem is a property of
        // a machine; a session only says where to start looking.
@@ -121,6 +137,15 @@ pub fn router(manager: Arc<SessionManager>) -> Router {
        .route("/sessions/{id}", get(read_session).delete(delete_session))
        .route("/sessions/{id}/events", get(events))
        .route("/sessions/{id}/transcript", get(transcript))
+        .route("/sessions/{id}/subagents", get(list_subagents))
+        .route(
+            "/sessions/{id}/subagents/{sub}/transcript",
+            get(subagent_transcript),
+        )
+        .route(
+            "/sessions/{id}/subagents/{sub}/events",
+            get(subagent_events),
+        )
        .route("/sessions/{id}/message", post(message))
        .route("/sessions/{id}/unqueue", post(unqueue))
        .route("/sessions/{id}/answer", post(answer))
@@ -131,7 +156,10 @@ pub fn router(manager: Arc<SessionManager>) -> Router {
        .route("/sessions/{id}/cwd", post(set_cwd))
        .route("/sessions/{id}/model", post(set_model))
        .route("/sessions/{id}/permission-mode", post(set_permission_mode))
+        .route("/sessions/{id}/effort", post(set_effort))
+        .route("/defaults", get(defaults).post(set_defaults))
        .route("/sessions/{id}/notify", post(set_notify))
+        .route("/sessions/{id}/auto-resume", post(set_auto_resume))
        .route("/notifications", get(notifications))
        .route("/sessions/{id}/compact", post(compact))
        .route("/sessions/{id}/command", post(command))
@@ -197,6 +225,18 @@ fn lookup(manager: &SessionManager, id: &str) -> Result<Arc<LiveSession>, ApiErr
        .ok_or_else(|| ApiError::NotFound(format!("no session {id}")))
 }

+/// A session's subagent by id -- the second half of the lookup every
+/// `/sessions/{id}/subagents/{sub}/...` route needs. `Arc` because reopening
+/// one from disk (a subagent this process has not touched yet) inserts it
+/// into the registry, and a route holding a borrow across that would be
+/// holding the registry's lock the whole request.
+fn lookup_subagent(session: &LiveSession, sub: &str) -> Result<Arc<Subagent>, ApiError> {
+    session
+        .subagents()
+        .get(sub)
+        .ok_or_else(|| ApiError::NotFound(format!("no subagent {sub}")))
+}
+
 async fn list_sessions(State(manager): State<Arc<SessionManager>>) -> axum::Json<Vec<SessionInfo>> {
    axum::Json(manager.sessions())
 }
@@ -291,6 +331,9 @@ struct SshRequest {
    /// Where attached files land on that machine; see `SshConfig`.
    #[serde(default)]
    attachments_dir: Option<String>,
+    /// Where that machine keeps its GGUF models; see `SshConfig`.
+    #[serde(default)]
+    models_dir: Option<String>,
 }

 impl SshRequest {
@@ -320,6 +363,14 @@ impl SshRequest {
                .map(str::trim)
                .filter(|dir| !dir.is_empty())
                .map(std::path::PathBuf::from),
+            // The same rule, and for the same reason: this directory is
+            // on the other machine, so a `~` in it is that machine's home.
+            models_dir: self
+                .models_dir
+                .as_deref()
+                .map(str::trim)
+                .filter(|dir| !dir.is_empty())
+                .map(std::path::PathBuf::from),
        })
    }
 }
@@ -495,6 +546,27 @@ struct PathQuery {
    path: String,
 }

+/// The models **that machine** has, which is the list a llama.cpp session
+/// on it can choose from.
+///
+/// Not `GET /models`, which is this backend's own downloads: those are on
+/// the machine a session runs on only when they are the same machine. A
+/// spawn screen offering this backend's list for a remote setup would be
+/// naming files that are not there, and the session would fail at the
+/// point of loading rather than at the point of choosing.
+async fn setup_models(
+    State(manager): State<Arc<SessionManager>>,
+    UrlPath(id): UrlPath<String>,
+) -> Result<axum::Json<Vec<crate::models::LocalModel>>, ApiError> {
+    let setup = setup_by_id(&manager, &id)?;
+    let transport = crate::session::transport::Transport::for_setup(&setup);
+    let dir = crate::models::dir_on(&transport, manager.models_dir());
+    crate::models::on_machine(&transport, &dir)
+        .await
+        .map(axum::Json)
+        .map_err(from_machine)
+}
+
 /// What is in a directory, and what that directory resolved to.
 async fn list_dir(
    State(manager): State<Arc<SessionManager>>,
@@ -609,6 +681,8 @@ struct SpawnRequest {
    cwd: Option<PathBuf>,
    #[serde(default)]
    permission_mode: Option<String>,
+    #[serde(default)]
+    effort: Option<String>,
    /// Whatever the chosen driver understands -- llama.cpp's context size and
    /// sampling. Opaque here on purpose: see `SessionConfig::params`.
    #[serde(default)]
@@ -796,6 +870,7 @@ async fn start_import(
            model: body.model.clone(),
            cwd: None,
            permission_mode: body.permission_mode.clone(),
+            effort: body.effort.clone(),
            params: std::collections::BTreeMap::new(),
            import: Some(session.clone()),
        };
@@ -828,6 +903,8 @@ struct ImportRequest {
    model: Option<String>,
    #[serde(default)]
    permission_mode: Option<String>,
+    #[serde(default)]
+    effort: Option<String>,
 }

 /// Runs `work` on the server, marked as in flight for as long as it takes.
@@ -978,6 +1055,7 @@ async fn spawn(manager: &Arc<SessionManager>, body: SpawnRequest) -> Result<Sess
                .filter(|cwd| cwd.as_os_str() != "")
        }),
        permission_mode: body.permission_mode,
+        effort: body.effort,
        params: body.params,
    };

@@ -1311,6 +1389,60 @@ struct PermissionModeRequest {
    mode: String,
 }

+/// What new sessions start at. One field today; a struct rather than a bare
+/// value because "the defaults" is the thing a phone asks for, and the next
+/// one to move here -- the permission mode, which the spawn screen still
+/// hardcodes -- must not need a second route.
+#[derive(Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+#[serde(deny_unknown_fields)]
+struct Defaults {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    effort: Option<String>,
+}
+
+async fn defaults(State(manager): State<Arc<SessionManager>>) -> axum::Json<Defaults> {
+    axum::Json(Defaults {
+        effort: manager.default_effort(),
+    })
+}
+
+/// Sets what a new session's thinking level is. Applied when a session is
+/// spawned, so nothing already running changes underneath anybody.
+async fn set_defaults(
+    State(manager): State<Arc<SessionManager>>,
+    axum::Json(body): axum::Json<Defaults>,
+) -> Result<StatusCode, ApiError> {
+    manager
+        .set_default_effort(body.effort.as_deref())
+        .map_err(bad_request)?;
+    Ok(StatusCode::NO_CONTENT)
+}
+
+#[derive(Deserialize)]
+#[serde(rename_all = "camelCase")]
+#[serde(deny_unknown_fields)]
+struct EffortRequest {
+    /// Absent or null is the CLI's own default, which is a choice somebody can
+    /// make rather than only a state to start in.
+    #[serde(default)]
+    effort: Option<String>,
+}
+
+/// Records how hard this session thinks, and stops the process so the next one
+/// is launched with it -- `--effort` has no control request behind it. See
+/// [`SessionManager::set_session_effort`].
+async fn set_effort(
+    State(manager): State<Arc<SessionManager>>,
+    UrlPath(id): UrlPath<String>,
+    axum::Json(body): axum::Json<EffortRequest>,
+) -> Result<StatusCode, ApiError> {
+    manager
+        .set_session_effort(&id, body.effort.as_deref())
+        .map_err(bad_request)?;
+    Ok(StatusCode::NO_CONTENT)
+}
+
 async fn set_permission_mode(
    State(manager): State<Arc<SessionManager>>,
    UrlPath(id): UrlPath<String>,
@@ -1339,6 +1471,33 @@ async fn set_notify(
    Ok(StatusCode::NO_CONTENT)
 }

+#[derive(Deserialize)]
+#[serde(rename_all = "camelCase", deny_unknown_fields)]
+struct AutoResumeRequest {
+    auto_resume: bool,
+    /// What to send when the limit lifts. Absent -- and empty, which is what a
+    /// cleared field sends -- means this app's own default word, which is a
+    /// choice a caller has to be able to make rather than only start in.
+    #[serde(default)]
+    message: Option<String>,
+}
+
+/// Turns auto-resume on or off, and sets what it would say.
+///
+/// One request for both, because they are one decision: switching it on
+/// without saying what to send is the ordinary case, and changing the words
+/// while it is off is how somebody sets it up before it is needed.
+async fn set_auto_resume(
+    State(manager): State<Arc<SessionManager>>,
+    UrlPath(id): UrlPath<String>,
+    axum::Json(body): axum::Json<AutoResumeRequest>,
+) -> Result<StatusCode, ApiError> {
+    manager
+        .set_session_auto_resume(&id, body.auto_resume, body.message.as_deref())
+        .map_err(bad_request)?;
+    Ok(StatusCode::NO_CONTENT)
+}
+
 #[derive(Deserialize)]
 #[serde(deny_unknown_fields)]
 struct CommandRequest {
@@ -1616,8 +1775,36 @@ async fn transcript(
    Query(query): Query<TranscriptQuery>,
 ) -> Result<axum::Json<Vec<crate::session::transcript::SeqEvent>>, ApiError> {
    let session = lookup(&manager, &id)?;
+    transcript_page(session.transcript_path(), &id, query)
+}
+
+/// Exactly [`transcript`]'s route and answer, against one subagent's own
+/// transcript instead of its session's -- see `SUBAGENTS.md`'s wire shape.
+async fn subagent_transcript(
+    State(manager): State<Arc<SessionManager>>,
+    UrlPath((id, sub)): UrlPath<(String, String)>,
+    Query(query): Query<TranscriptQuery>,
+) -> Result<axum::Json<Vec<crate::session::transcript::SeqEvent>>, ApiError> {
+    let session = lookup(&manager, &id)?;
+    let subagent = lookup_subagent(&session, &sub)?;
+    transcript_page(
+        &subagent.transcript_path(),
+        &format!("{id}/subagents/{sub}"),
+        query,
+    )
+}
+
+/// A page of history at `path`, newest first to open with -- the one
+/// implementation [`transcript`] and [`subagent_transcript`] share, since a
+/// subagent's transcript is read exactly the way a session's is. `label` is
+/// only for the debug line below.
+fn transcript_page(
+    path: &Path,
+    label: &str,
+    query: TranscriptQuery,
+) -> Result<axum::Json<Vec<crate::session::transcript::SeqEvent>>, ApiError> {
    let events = crate::session::transcript::read_window(
-        session.transcript_path(),
+        path,
        query.before,
        query.after,
        query.limit,
@@ -1629,7 +1816,7 @@ async fn transcript(
    // for events and draws rows, and the ratio between them is a property of
    // the conversation. `RUST_LOG=ai_server=debug`.
    tracing::debug!(
-        session = %id,
+        session = %label,
        before = ?query.before,
        after = ?query.after,
        limit = query.limit,
@@ -1647,23 +1834,74 @@ async fn events(
    headers: HeaderMap,
 ) -> Result<Sse<impl tokio_stream::Stream<Item = Result<SseEvent, Infallible>>>, ApiError> {
    let session = lookup(&manager, &id)?;
-    let cursor = headers
-        .get("last-event-id")
-        .and_then(|value| value.to_str().ok())
-        .and_then(|value| value.parse().ok())
-        .unwrap_or(query.after);
-
+    let cursor = cursor_of(&headers, query.after);
    // Subscribe before reading the file so nothing can land in the gap
    // between replay and live; overlap is deduplicated by seq.
    let live = session.subscribe();
-    let (tx, stream) = mpsc::channel(64);
-    tokio::spawn(stream_session(
+    Ok(sse_stream(
        session.transcript_path().to_path_buf(),
        cursor,
        live,
-        tx,
-    ));
-    Ok(Sse::new(ReceiverStream::new(stream).map(Ok)).keep_alive(KeepAlive::default()))
+    ))
+}
+
+/// Exactly [`events`]'s route and answer, against one subagent's own stream
+/// instead of its session's -- see `SUBAGENTS.md`'s wire shape.
+async fn subagent_events(
+    State(manager): State<Arc<SessionManager>>,
+    UrlPath((id, sub)): UrlPath<(String, String)>,
+    Query(query): Query<EventsQuery>,
+    headers: HeaderMap,
+) -> Result<Sse<impl tokio_stream::Stream<Item = Result<SseEvent, Infallible>>>, ApiError> {
+    let session = lookup(&manager, &id)?;
+    let subagent = lookup_subagent(&session, &sub)?;
+    let cursor = cursor_of(&headers, query.after);
+    let live = subagent.subscribe();
+    Ok(sse_stream(subagent.transcript_path(), cursor, live))
+}
+
+/// The cursor an SSE reconnect resumes from: the native `Last-Event-ID`
+/// takes precedence over the query parameter, same cursor either way.
+fn cursor_of(headers: &HeaderMap, query_after: u64) -> u64 {
+    headers
+        .get("last-event-id")
+        .and_then(|value| value.to_str().ok())
+        .and_then(|value| value.parse().ok())
+        .unwrap_or(query_after)
+}
+
+/// Spawns the backlog-then-live task and wraps it as the response, the one
+/// piece [`events`] and [`subagent_events`] share.
+fn sse_stream(
+    transcript: PathBuf,
+    cursor: u64,
+    live: broadcast::Receiver<SeqEvent>,
+) -> Sse<impl tokio_stream::Stream<Item = Result<SseEvent, Infallible>>> {
+    let (tx, stream) = mpsc::channel(64);
+    tokio::spawn(stream_session(transcript, cursor, live, tx));
+    Sse::new(ReceiverStream::new(stream).map(Ok)).keep_alive(KeepAlive::default())
+}
+
+/// `GET /sessions/{id}/subagents`: every subagent this session has started,
+/// oldest first, with a status read from its own transcript -- see
+/// `SUBAGENTS.md`'s wire shape. A subagent whose last status is `Running` is
+/// reported `Unknown` instead when the session itself is not running: its
+/// process was the session's, and a session with none has nothing left to
+/// ask.
+async fn list_subagents(
+    State(manager): State<Arc<SessionManager>>,
+    UrlPath(id): UrlPath<String>,
+) -> Result<axum::Json<Vec<SubagentInfo>>, ApiError> {
+    let session = lookup(&manager, &id)?;
+    // Anything but `Exited` or `Unknown` has a process behind it, which is
+    // what decides whether a subagent still reading `Running` from its own
+    // transcript can be believed -- see `SUBAGENTS.md`'s wire shape.
+    let running = !matches!(
+        session.status(),
+        crate::session::driver::SessionStatus::Exited
+            | crate::session::driver::SessionStatus::Unknown
+    );
+    Ok(axum::Json(session.subagents().list(running)))
 }

 /// Every session's attention-wanting moments, on one stream.
@@ -59,6 +59,7 @@ use tokio::sync::mpsc;

 use super::driver::{AttachmentRef, Driver, Event, EventSink, SessionStatus, Unqueued};
 use super::process;
+use super::subagent::Subagents;
 use super::transport::{Launch, Streams, Transport};
 use crate::config::{ProviderConfig, SessionConfig};
 use translate::{AnswerOutcome, Setting, Translator, starts_a_model_call};
@@ -213,8 +214,12 @@ impl ClaudeDriver {
        transport: &Transport,
        session_dir: &Path,
        sink: EventSink,
+        subagents: Arc<Subagents>,
    ) -> Result<Self> {
-        let state = Arc::new(Mutex::new(Translator::new(session_dir.to_path_buf())));
+        let state = Arc::new(Mutex::new(Translator::new(
+            session_dir.to_path_buf(),
+            subagents,
+        )));
        let queue = Arc::new(Mutex::new(Queue::default()));
        let reading = Arc::new(AtomicBool::new(true));

@@ -351,6 +356,12 @@ impl ClaudeDriver {
        if let Some(mode) = &meta.permission_mode {
            push("--permission-mode", mode);
        }
+        // Launch-only: see `SessionConfig::effort`. Omitted entirely when
+        // unset, so the CLI's own default is what an unchosen session gets
+        // rather than a level this app decided to call the default.
+        if let Some(effort) = &meta.effort {
+            push("--effort", effort);
+        }
        // Named at birth, so this session is the same session in the CLI's own
        // picker and in what other agents see.
        //
@@ -386,7 +397,7 @@ impl ClaudeDriver {
        let stdout = create_log(&session_dir.join(STDOUT_LOG))?;
        let stderr = create_log(&session_dir.join(STDERR_LOG))?;

-        let program = provider.command.as_deref().unwrap_or("claude");
+        let program = provider.program();
        let launch = Launch::new(program, args, meta.cwd.as_deref());
        let child = transport.spawn(
            &launch,
@@ -1163,7 +1174,10 @@ mod tests {
    /// this" and "the transcript records that".
    fn events_from_lines(lines: &[&str]) -> Vec<Event> {
        let dir = tempfile::tempdir().expect("temp dir");
-        let state = Arc::new(Mutex::new(Translator::new(dir.path().to_path_buf())));
+        let state = Arc::new(Mutex::new(Translator::new(
+            dir.path().to_path_buf(),
+            Arc::new(Subagents::new(dir.path().to_path_buf())),
+        )));
        let queue = Arc::new(Mutex::new(Queue::default()));
        let (sink, mut out) = mpsc::unbounded_channel::<Event>();
        for line in lines {
@@ -1187,7 +1201,10 @@ mod tests {
        interject: impl FnOnce(&Arc<Mutex<Queue>>),
    ) -> Vec<Event> {
        let dir = tempfile::tempdir().expect("temp dir");
-        let state = Arc::new(Mutex::new(Translator::new(dir.path().to_path_buf())));
+        let state = Arc::new(Mutex::new(Translator::new(
+            dir.path().to_path_buf(),
+            Arc::new(Subagents::new(dir.path().to_path_buf())),
+        )));
        let queue = Arc::new(Mutex::new(Queue::default()));
        let (sink, mut out) = mpsc::unbounded_channel::<Event>();
        let mut interject = Some(interject);
@@ -1481,7 +1498,10 @@ mod tests {
        // doing.
        let dir = tempfile::tempdir().expect("tempdir");
        let (sink, mut received) = mpsc::unbounded_channel();
-        let state = Arc::new(Mutex::new(Translator::new(dir.path().to_path_buf())));
+        let state = Arc::new(Mutex::new(Translator::new(
+            dir.path().to_path_buf(),
+            Arc::new(Subagents::new(dir.path().to_path_buf())),
+        )));
        let queue = Arc::new(Mutex::new(Queue::default()));

        let text = r#"{"type":"stream_event","event":{"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"working"}},"parent_tool_use_id":null}"#;
@@ -1526,7 +1546,10 @@ mod tests {
        // session back to work.
        let dir = tempfile::tempdir().expect("tempdir");
        let (sink, mut received) = mpsc::unbounded_channel();
-        let state = Arc::new(Mutex::new(Translator::new(dir.path().to_path_buf())));
+        let state = Arc::new(Mutex::new(Translator::new(
+            dir.path().to_path_buf(),
+            Arc::new(Subagents::new(dir.path().to_path_buf())),
+        )));
        let queue = Arc::new(Mutex::new(Queue::default()));
        queue.lock().unwrap().close(&sink, "the session ended");

@@ -11,10 +11,12 @@

 use std::collections::HashMap;
 use std::path::{Path, PathBuf};
+use std::sync::{Arc, Mutex};

 use serde_json::{Value, json};

 use super::super::driver::{Event, QuestionOption, SessionStatus, context_tokens};
+use super::super::subagent::Subagents;

 /// Whether this line is the CLI opening a fresh model call.
 ///
@@ -92,10 +94,20 @@ pub(super) struct Translator {
    /// reports none rather than repeating the previous turn's.
    context: Option<u64>,
    session_dir: PathBuf,
+    /// This session's subagents, shared with every child translator below --
+    /// see `SUBAGENTS.md`. One registry per session, so a subagent started
+    /// through this translator or any of its children lands in the same
+    /// place a route reads it back from.
+    subagents: Arc<Subagents>,
+    /// One translator per subagent id, holding *its* streaming and
+    /// tool-tracking state -- separate from the parent's because tool ids
+    /// are unique but a `stream_event`'s content-block index is not, and
+    /// parallel subagents interleave their deltas on one stdout.
+    children: HashMap<String, Arc<Mutex<Translator>>>,
 }

 impl Translator {
-    pub(super) fn new(session_dir: PathBuf) -> Self {
+    pub(super) fn new(session_dir: PathBuf, subagents: Arc<Subagents>) -> Self {
        Self {
            session_id: None,
            pending: HashMap::new(),
@@ -103,6 +115,8 @@ impl Translator {
            interrupting: false,
            context: None,
            session_dir,
+            subagents,
+            children: HashMap::new(),
        }
    }

@@ -122,13 +136,77 @@ impl Translator {
    pub(super) fn translate(&mut self, message: &Value) -> Vec<Event> {
        // Events from subagents (Task tool internals) carry a
        // parent_tool_use_id; the transcript shows the Task tool's own
-        // start/end instead of every nested step.
-        if message
-            .get("parent_tool_use_id")
-            .is_some_and(|id| !id.is_null())
-        {
-            return Vec::new();
+        // start/end instead of every nested step. Routed into that
+        // subagent's own transcript rather than dropped -- see
+        // `SUBAGENTS.md`.
+        if let Some(parent_id) = message.get("parent_tool_use_id").and_then(Value::as_str) {
+            return self.translate_child(parent_id, message);
        }
+        self.dispatch(message)
+    }
+
+    /// A line belonging to a subagent rather than to this translator's own
+    /// session. Always returns nothing to the *caller*: everything it
+    /// produces goes into the subagent's own transcript instead.
+    fn translate_child(&mut self, id: &str, message: &Value) -> Vec<Event> {
+        match self.subagents.get(id) {
+            Some(subagent) if !subagent.is_open() => {
+                // Not stale: the Task tool runs in the background by
+                // default, so a finished subagent can still be sent another
+                // message later (SendMessage) and start working again. A
+                // line arriving after `finish` means exactly that, not a
+                // conversation that is over -- see `SUBAGENTS.md`.
+                self.subagents.reopen(id);
+            }
+            Some(_) => {}
+            None => {
+                // Nobody has heard of this id yet: the Task call itself
+                // either has not been seen or never will be. Started here
+                // with the best title available -- the tool name of this
+                // first line -- since SUBAGENTS.md's real title only
+                // arrives with the Task call.
+                self.subagents.start(id, &fallback_title(message), None);
+            }
+        }
+        let child = self
+            .children
+            .entry(id.to_string())
+            .or_insert_with(|| {
+                Arc::new(Mutex::new(Translator::new(
+                    self.session_dir.clone(),
+                    Arc::clone(&self.subagents),
+                )))
+            })
+            .clone();
+        let events = child.lock().unwrap().dispatch(message);
+        for event in events {
+            // The subagent's own vocabulary is Running/Exited/Unknown, never
+            // Idle -- a background Task is either working or it has ended,
+            // never merely "between turns" the way a session is. Dropped
+            // here rather than never produced, so a `result` line's own
+            // `Idle` (dispatch's ordinary end-of-turn event, for a subagent
+            // dialect that ever sends one) is caught the same way a
+            // `message_delta` would be.
+            if !matches!(
+                event,
+                Event::Status {
+                    state: SessionStatus::Idle
+                }
+            ) {
+                self.subagents.record(id, event);
+            }
+        }
+        // What actually ends a subagent's turn: not the parent's
+        // `tool_result`, which for a background Task arrives at launch
+        // ("Async agent launched...") long before the work is done -- see
+        // `SUBAGENTS.md`.
+        if ends_a_turn(message) {
+            self.subagents.finish(id);
+        }
+        Vec::new()
+    }
+
+    fn dispatch(&mut self, message: &Value) -> Vec<Event> {
        match message.get("type").and_then(Value::as_str) {
            Some("system") => self.translate_system(message),
            // The CLI's own announcement that `/clear` took effect, sent just
@@ -224,12 +302,12 @@ impl Translator {
                        .and_then(Value::as_bool)
                        .unwrap_or(false)
                {
-                    events.push(Event::Error {
-                        message: message
-                            .get("result")
-                            .and_then(Value::as_str)
-                            .unwrap_or("the turn ended with an error")
-                            .to_string(),
+                    let said = message.get("result").and_then(Value::as_str);
+                    events.push(match said.and_then(usage_limit) {
+                        Some(resets_at) => Event::LimitReached { resets_at },
+                        None => Event::Error {
+                            message: said.unwrap_or("the turn ended with an error").to_string(),
+                        },
                    });
                }
                let context = self.context.take();
@@ -379,22 +457,47 @@ impl Translator {
        content
            .iter()
            .filter(|block| block.get("type").and_then(Value::as_str) == Some("tool_use"))
-            .map(|block| Event::ToolStart {
-                id: block
+            .map(|block| {
+                let id = block
                    .get("id")
                    .and_then(Value::as_str)
                    .unwrap_or_default()
-                    .to_string(),
-                tool: block
+                    .to_string();
+                let tool = block
                    .get("name")
                    .and_then(Value::as_str)
                    .unwrap_or_default()
-                    .to_string(),
-                input: block.get("input").cloned().unwrap_or(Value::Null),
+                    .to_string();
+                let input = block.get("input").cloned().unwrap_or(Value::Null);
+                // A subagent this call is about to start -- see
+                // `SUBAGENTS.md`'s lifecycle #1. The parent's own transcript
+                // still shows only the Task call itself, below.
+                if tool == "Task" || tool == "Agent" {
+                    self.start_subagent_from_task(&id, &input);
+                }
+                Event::ToolStart { id, tool, input }
            })
            .collect()
    }

+    /// Starts the subagent a Task call names, with the title and prompt
+    /// SUBAGENTS.md describes: the call's `description`, then
+    /// `(<subagent_type>)` when one is given, falling back to the tool's own
+    /// name when there is no description to build one from.
+    fn start_subagent_from_task(&self, id: &str, input: &Value) {
+        let description = text_field(input, "description");
+        let subagent_type = text_field(input, "subagent_type");
+        let prompt = input.get("prompt").and_then(Value::as_str);
+        let title = match (description, subagent_type) {
+            (Some(description), Some(subagent_type)) => {
+                format!("{description} ({subagent_type})")
+            }
+            (Some(description), None) => description,
+            (None, _) => "Task".to_string(),
+        };
+        self.subagents.start(id, &title, prompt);
+    }
+
    fn translate_control_request(&mut self, message: &Value) -> Vec<Event> {
        let request = &message["request"];
        if request.get("subtype").and_then(Value::as_str) != Some("can_use_tool") {
@@ -598,11 +701,90 @@ impl Translator {
                id: about.clone(),
                output: texts.join("\n"),
            });
+            // Deliberately does *not* finish a subagent `about` might name:
+            // the Task tool runs in the background by default, so this
+            // `tool_result` -- "Async agent launched..." -- arrives at
+            // launch, long before the subagent's own work is done. What
+            // ends it is its own turn ending, handled in `translate_child`.
        }
        events
    }
 }

+/// The title to start a subagent under when its own first line arrives
+/// before (or without) its Task call ever being seen: the tool name of that
+/// first line, which is the only thing known about it yet. `"subagent"` for
+/// a line this cannot even find a tool name in, such as one that opens with
+/// something other than a tool call.
+fn fallback_title(message: &Value) -> String {
+    message["message"]["content"]
+        .as_array()
+        .into_iter()
+        .flatten()
+        .find(|block| block.get("type").and_then(Value::as_str) == Some("tool_use"))
+        .and_then(|block| block.get("name"))
+        .and_then(Value::as_str)
+        .unwrap_or("subagent")
+        .to_string()
+}
+
+/// Whether this line is a subagent's *own* turn ending -- the only thing
+/// that does, per `SUBAGENTS.md`: not the parent's `tool_result`, which for
+/// a background Task arrives at launch rather than at completion.
+///
+/// Checked on the raw line rather than on what `dispatch` returns, so this
+/// never has to touch the shared `translate_stream_event`/`dispatch` code a
+/// top-level session's own turn-ending also goes through -- a subagent's
+/// idea of "ended" must not change when a real session's does.
+///
+/// `message_delta` is the raw API's own signal, carrying the stop reason:
+/// `end_turn` is genuinely done, `tool_use` means the model is about to call
+/// one and there is more coming. A `result` line is the CLI's own shape for
+/// a top-level turn; a subagent has not been observed to send one, but
+/// SUBAGENTS.md counts it too in case a future CLI version does.
+fn ends_a_turn(message: &Value) -> bool {
+    match message.get("type").and_then(Value::as_str) {
+        Some("stream_event") => {
+            let event = &message["event"];
+            event.get("type").and_then(Value::as_str) == Some("message_delta")
+                && event["delta"].get("stop_reason").and_then(Value::as_str) == Some("end_turn")
+        }
+        Some("result") => true,
+        _ => false,
+    }
+}
+
+/// Whether a failed turn failed because the account is out of quota, and when
+/// the CLI said the limit lifts.
+///
+/// The wording is the CLI's: a turn stopped by the limit ends with `is_error`
+/// and a result of `Claude AI usage limit reached|1788546972`, the reset being
+/// epoch seconds after a pipe. Matched on the sentence rather than on a code
+/// because the CLI sends none, so this is deliberately loose about everything
+/// but the four words.
+///
+/// The two `None`s mean different things and both are real. The outer one is
+/// "some other failure". The inner one is "the limit is reached and the CLI did
+/// not say until when" -- which is not a reason to invent a time: `crate::resume`
+/// asks the usage endpoint before sending anything, and that answer is the one
+/// that decides.
+///
+/// Milliseconds are accepted as well as seconds and told apart by magnitude,
+/// since a wrong guess would schedule a resume tens of thousands of years out
+/// and look exactly like auto-resume being broken.
+fn usage_limit(result: &str) -> Option<Option<f64>> {
+    if !result.to_ascii_lowercase().contains("usage limit reached") {
+        return None;
+    }
+    let stamp = result
+        .rsplit('|')
+        .next()
+        .and_then(|tail| tail.trim().parse::<f64>().ok())
+        .filter(|stamp| *stamp > 0.0)
+        .map(|stamp| if stamp > 1e11 { stamp / 1000.0 } else { stamp });
+    Some(stamp)
+}
+
 /// A string field that is there and not empty, or `None`. The CLI omits these
 /// rather than sending them empty, but a caller that sends `""` means the same
 /// thing and should not produce a description that draws as a blank line.
@@ -665,10 +847,18 @@ mod tests {
            .collect()
    }

+    /// A fresh, empty subagent registry over the same temp dir a test's
+    /// translator writes into -- every test here is about the parent's own
+    /// events, so what a registry does with a subagent is `subagent.rs`'s
+    /// tests to make, not these.
+    fn test_subagents(dir: &tempfile::TempDir) -> Arc<Subagents> {
+        Arc::new(Subagents::new(dir.path().to_path_buf()))
+    }
+
    #[test]
    fn captures_the_resume_token_and_the_settings_from_init() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -690,7 +880,7 @@ mod tests {
    #[test]
    fn a_setting_is_reported_when_the_cli_accepts_it_and_not_before() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));

        // What `set_model` does: remember, send, and say nothing yet.
        translator.expect_setting("req-a".to_string(), Setting::Model("sonnet".to_string()));
@@ -767,7 +957,7 @@ mod tests {
        // The line it sends just after answering `set_permission_mode`, which is
        // also how a mode changed from the terminal arrives.
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -787,7 +977,7 @@ mod tests {
    fn streams_text_deltas_and_skips_the_consolidated_copy() {
        // Real lines (trimmed) from the 2.1.237 probe.
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -807,7 +997,7 @@ mod tests {
    #[test]
    fn tool_use_and_result_become_tool_events() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -834,7 +1024,7 @@ mod tests {
    #[test]
    fn subagent_events_are_not_duplicated_into_the_transcript() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -844,10 +1034,243 @@ mod tests {
        assert!(events.is_empty());
    }

+    /// A child line does not just vanish from the parent -- it lands in its
+    /// own subagent's transcript, with that transcript's own sequence
+    /// numbers, starting at 1 like any other.
+    #[test]
+    fn a_child_line_lands_in_its_own_subagents_transcript() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let subagents = test_subagents(&dir);
+        let mut translator = Translator::new(dir.path().to_path_buf(), Arc::clone(&subagents));
+        translate_lines(
+            &mut translator,
+            &[
+                r#"{"type":"assistant","message":{"content":[{"type":"tool_use","id":"toolu_c1","name":"Bash","input":{"command":"echo hi"}}]},"parent_tool_use_id":"toolu_parent"}"#,
+            ],
+        );
+        let subagent = subagents.get("toolu_parent").expect("subagent started");
+        let lines = crate::session::transcript::read_after(&subagent.transcript_path(), 0)
+            .expect("read subagent transcript");
+        assert_eq!(lines[0].seq, 1);
+        assert_eq!(
+            lines[0].event,
+            Event::Status {
+                state: SessionStatus::Running
+            }
+        );
+        assert!(
+            lines.iter().any(
+                |entry| matches!(&entry.event, Event::ToolStart { tool, .. } if tool == "Bash")
+            )
+        );
+    }
+
+    /// The title and prompt shown for a subagent come from the Task call
+    /// that started it, not from anything guessed at its first line.
+    #[test]
+    fn the_subagent_takes_its_title_and_prompt_from_the_task_call() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let subagents = test_subagents(&dir);
+        let mut translator = Translator::new(dir.path().to_path_buf(), Arc::clone(&subagents));
+        translate_lines(
+            &mut translator,
+            &[
+                r#"{"type":"assistant","message":{"content":[{"type":"tool_use","id":"toolu_task","name":"Task","input":{"description":"Investigate the bug","prompt":"Find why X fails","subagent_type":"general-purpose"}}]},"parent_tool_use_id":null}"#,
+            ],
+        );
+        let rows = subagents.list(true);
+        assert_eq!(rows.len(), 1);
+        assert_eq!(rows[0].title, "Investigate the bug (general-purpose)");
+        let subagent = subagents.get(&rows[0].id).expect("subagent");
+        let lines = crate::session::transcript::read_after(&subagent.transcript_path(), 0)
+            .expect("read subagent transcript");
+        assert!(lines.iter().any(
+            |entry| matches!(&entry.event, Event::UserMessage { text, .. } if text == "Find why X fails")
+        ));
+    }
+
+    /// The parent's `tool_result` for the Task id is what ends the
+    /// subagent -- SUBAGENTS.md's lifecycle #3 -- and nothing else does.
+    #[test]
+    fn the_parents_tool_result_does_not_finish_the_subagent() {
+        // The Task tool runs in the background by default: this
+        // `tool_result` is "Async agent launched...", arriving the moment
+        // the subagent *starts*, while it goes on working for however long
+        // its own turn takes. Finishing it here was the bug -- a running
+        // background agent read as "finished" with its transcript truncated
+        // at launch.
+        let dir = tempfile::tempdir().expect("tempdir");
+        let subagents = test_subagents(&dir);
+        let mut translator = Translator::new(dir.path().to_path_buf(), Arc::clone(&subagents));
+        translate_lines(
+            &mut translator,
+            &[
+                r#"{"type":"assistant","message":{"content":[{"type":"tool_use","id":"toolu_task2","name":"Task","input":{"description":"helper"}}]},"parent_tool_use_id":null}"#,
+            ],
+        );
+        let subagent = subagents.get("toolu_task2").expect("subagent started");
+        assert!(subagent.is_open());
+        translate_lines(
+            &mut translator,
+            &[
+                r#"{"type":"user","message":{"role":"user","content":[{"type":"tool_result","tool_use_id":"toolu_task2","content":"Async agent launched","is_error":false}]},"parent_tool_use_id":null}"#,
+            ],
+        );
+        assert!(subagent.is_open());
+    }
+
+    /// What actually ends a subagent: the raw API's own `message_delta`
+    /// saying its turn stopped with `end_turn`. Never written into the
+    /// subagent's own transcript as `Idle` -- its vocabulary has no such
+    /// state.
+    #[test]
+    fn the_subagents_own_end_turn_finishes_it() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let subagents = test_subagents(&dir);
+        let mut translator = Translator::new(dir.path().to_path_buf(), Arc::clone(&subagents));
+        translate_lines(
+            &mut translator,
+            &[
+                r#"{"type":"assistant","message":{"content":[{"type":"tool_use","id":"toolu_task3","name":"Task","input":{"description":"helper"}}]},"parent_tool_use_id":null}"#,
+                r#"{"type":"stream_event","event":{"type":"message_delta","delta":{"stop_reason":"end_turn"}},"parent_tool_use_id":"toolu_task3"}"#,
+            ],
+        );
+        let subagent = subagents.get("toolu_task3").expect("subagent started");
+        assert!(!subagent.is_open());
+        let lines = crate::session::transcript::read_after(&subagent.transcript_path(), 0)
+            .expect("read subagent transcript");
+        assert!(
+            !lines
+                .iter()
+                .any(|entry| matches!(&entry.event, Event::Status { state } if *state == SessionStatus::Idle)),
+            "a subagent's transcript must never carry Idle: {lines:?}"
+        );
+        assert_eq!(
+            lines.last().unwrap().event,
+            Event::Status {
+                state: SessionStatus::Exited
+            }
+        );
+    }
+
+    /// `stop_reason: "tool_use"` is the model about to call a tool, with
+    /// more of the turn still coming -- not an end.
+    #[test]
+    fn a_stop_reason_of_tool_use_does_not_finish_the_subagent() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let subagents = test_subagents(&dir);
+        let mut translator = Translator::new(dir.path().to_path_buf(), Arc::clone(&subagents));
+        translate_lines(
+            &mut translator,
+            &[
+                r#"{"type":"assistant","message":{"content":[{"type":"tool_use","id":"toolu_task4","name":"Task","input":{"description":"helper"}}]},"parent_tool_use_id":null}"#,
+                r#"{"type":"stream_event","event":{"type":"message_delta","delta":{"stop_reason":"tool_use"}},"parent_tool_use_id":"toolu_task4"}"#,
+            ],
+        );
+        assert!(
+            subagents
+                .get("toolu_task4")
+                .expect("subagent started")
+                .is_open()
+        );
+    }
+
+    /// A background Task can be sent another message long after its first
+    /// turn ended -- a further child line for it reopens rather than being
+    /// dropped, and the same transcript and child translator carry on.
+    #[test]
+    fn a_line_after_finish_reopens_the_subagent_rather_than_being_dropped() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let subagents = test_subagents(&dir);
+        let mut translator = Translator::new(dir.path().to_path_buf(), Arc::clone(&subagents));
+        translate_lines(
+            &mut translator,
+            &[
+                r#"{"type":"assistant","message":{"content":[{"type":"tool_use","id":"toolu_task5","name":"Task","input":{"description":"helper"}}]},"parent_tool_use_id":null}"#,
+                r#"{"type":"stream_event","event":{"type":"message_delta","delta":{"stop_reason":"end_turn"}},"parent_tool_use_id":"toolu_task5"}"#,
+            ],
+        );
+        let subagent = subagents.get("toolu_task5").expect("subagent started");
+        assert!(!subagent.is_open());
+
+        translate_lines(
+            &mut translator,
+            &[
+                r#"{"type":"assistant","message":{"content":[{"type":"tool_use","id":"toolu_more","name":"Bash","input":{}}]},"parent_tool_use_id":"toolu_task5"}"#,
+            ],
+        );
+        assert!(subagent.is_open());
+        let lines = crate::session::transcript::read_after(&subagent.transcript_path(), 0)
+            .expect("read subagent transcript");
+        // Running, [prompt], Exited, Running (reopened), then the new line's
+        // own ToolStart -- the same transcript throughout, not a new one.
+        assert!(
+            lines.iter().any(
+                |entry| matches!(&entry.event, Event::ToolStart { tool, .. } if tool == "Bash")
+            )
+        );
+        assert_eq!(
+            lines
+                .iter()
+                .filter(|entry| matches!(
+                    &entry.event,
+                    Event::Status {
+                        state: SessionStatus::Running
+                    }
+                ))
+                .count(),
+            2,
+            "expected one Running at creation and one at the reopen: {lines:?}"
+        );
+    }
+
+    /// Two subagents running at once keep two separate transcripts: tool ids
+    /// are unique but a `stream_event`'s content-block index is not, so
+    /// sharing translation state between them would cross their streams.
+    #[test]
+    fn two_parallel_subagents_keep_separate_transcripts() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let subagents = test_subagents(&dir);
+        let mut translator = Translator::new(dir.path().to_path_buf(), Arc::clone(&subagents));
+        translate_lines(
+            &mut translator,
+            &[
+                r#"{"type":"assistant","message":{"content":[{"type":"tool_use","id":"toolu_a","name":"Bash","input":{}}]},"parent_tool_use_id":"toolu_task_a"}"#,
+                r#"{"type":"assistant","message":{"content":[{"type":"tool_use","id":"toolu_b","name":"Read","input":{}}]},"parent_tool_use_id":"toolu_task_b"}"#,
+            ],
+        );
+        let a = subagents.get("toolu_task_a").expect("subagent a");
+        let b = subagents.get("toolu_task_b").expect("subagent b");
+        let a_events = crate::session::transcript::read_after(&a.transcript_path(), 0)
+            .expect("read a's transcript");
+        let b_events = crate::session::transcript::read_after(&b.transcript_path(), 0)
+            .expect("read b's transcript");
+        assert!(
+            a_events.iter().any(
+                |entry| matches!(&entry.event, Event::ToolStart { tool, .. } if tool == "Bash")
+            )
+        );
+        assert!(
+            b_events.iter().any(
+                |entry| matches!(&entry.event, Event::ToolStart { tool, .. } if tool == "Read")
+            )
+        );
+        assert!(
+            !a_events.iter().any(
+                |entry| matches!(&entry.event, Event::ToolStart { tool, .. } if tool == "Read")
+            )
+        );
+        assert!(
+            !b_events.iter().any(
+                |entry| matches!(&entry.event, Event::ToolStart { tool, .. } if tool == "Bash")
+            )
+        );
+    }
+
    #[test]
    fn a_permission_request_becomes_an_allow_deny_question() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -896,7 +1319,7 @@ mod tests {
    #[test]
    fn denying_a_permission_sends_deny() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        translate_lines(
            &mut translator,
            &[
@@ -914,7 +1337,7 @@ mod tests {
        // The real 2.1.237 shape, verified live: answers go back inside
        // updatedInput, keyed by the question text.
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -966,7 +1389,7 @@ mod tests {
        // in the event: a phone that had to read this dialect's tool input to
        // find them would be the only place that knew how.
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -1014,7 +1437,7 @@ mod tests {
    #[test]
    fn images_in_tool_results_are_saved_and_referenced() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        // A 1x1 PNG, the smallest real payload worth round-tripping.
        let png = "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNk+M9QDwADhgGAWjR9awAAAABJRU5ErkJggg==";
        let line = format!(
@@ -1043,7 +1466,7 @@ mod tests {
    #[test]
    fn a_turn_result_reports_usage_and_returns_to_idle() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -1078,7 +1501,7 @@ mod tests {
    #[test]
    fn a_turn_started_by_another_agent_records_who_and_what() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -1112,7 +1535,7 @@ mod tests {
    #[test]
    fn an_ordinary_turn_carries_no_peer_note() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -1137,7 +1560,7 @@ mod tests {
    #[test]
    fn the_context_is_what_the_last_message_held_not_the_turn_added_up() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -1177,7 +1600,7 @@ mod tests {
        // Note the snake_case keys -- the CLI's transcript file writes the same
        // records in camelCase.
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -1207,7 +1630,7 @@ mod tests {
    #[test]
    fn a_failed_compaction_says_why_and_leaves_the_turn_running() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -1234,7 +1657,7 @@ mod tests {
    #[test]
    fn a_boundary_without_counts_says_so_rather_than_inventing_them() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -1254,7 +1677,7 @@ mod tests {
    #[test]
    fn an_error_result_surfaces_the_message() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -1275,6 +1698,55 @@ mod tests {
        );
    }

+    /// Running out of quota is a state, not a failure of the work.
+    ///
+    /// The naive reading -- an error result like any other -- is what shipped
+    /// before this: the transcript said "Claude AI usage limit reached|…" in
+    /// red, which is neither readable nor actionable, and nothing above the
+    /// driver could tell it apart from a broken tool call.
+    #[test]
+    fn a_turn_stopped_by_the_usage_limit_says_so_and_carries_the_reset() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
+        let events = translate_lines(
+            &mut translator,
+            &[
+                r#"{"type":"result","subtype":"error_during_execution","is_error":true,"result":"Claude AI usage limit reached|1788546972","usage":{}}"#,
+            ],
+        );
+        assert_eq!(
+            events[0],
+            Event::LimitReached {
+                resets_at: Some(1_788_546_972.0)
+            }
+        );
+    }
+
+    #[test]
+    fn a_limit_the_cli_gave_no_reset_for_is_reported_without_one() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
+        let events = translate_lines(
+            &mut translator,
+            &[
+                r#"{"type":"result","subtype":"error_during_execution","is_error":true,"result":"Claude AI usage limit reached","usage":{}}"#,
+            ],
+        );
+        // Not a time this side invented: the meter is asked before anything is
+        // sent, and a made-up reset would only decide when to ask.
+        assert_eq!(events[0], Event::LimitReached { resets_at: None });
+    }
+
+    #[test]
+    fn a_reset_in_milliseconds_is_not_read_as_the_year_58000() {
+        assert_eq!(
+            usage_limit("Claude AI usage limit reached|1788546972000"),
+            Some(Some(1_788_546_972.0))
+        );
+        // And anything that is not the limit stays an ordinary failure.
+        assert_eq!(usage_limit("something broke"), None);
+    }
+
    /// Pressing Stop is not a failure, and the CLI cannot tell you which it was.
    ///
    /// An interrupted turn arrives as exactly the same shape a broken one does,
@@ -1286,7 +1758,7 @@ mod tests {
    #[test]
    fn a_turn_stopped_on_purpose_is_not_an_error() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let stopped_result = r#"{"type":"result","subtype":"error_during_execution","is_error":true,"result":"Interrupted by user","usage":{}}"#;

        translator.expect_interrupt();
@@ -1318,7 +1790,7 @@ mod tests {
    #[test]
    fn replayed_and_synthetic_user_text_is_skipped() {
        let dir = tempfile::tempdir().expect("tempdir");
-        let mut translator = Translator::new(dir.path().to_path_buf());
+        let mut translator = Translator::new(dir.path().to_path_buf(), test_subagents(&dir));
        let events = translate_lines(
            &mut translator,
            &[
@@ -26,6 +26,20 @@
 //! - `/error [text]` -- a failure, which is otherwise awkward to cause.
 //! - `/peer [text]`, `/peer-turn` -- a message from another agent, in the
 //!   in-place and the live shapes.
+//! - `/usage [what]` -- an invented rate-limit answer, or `/usage off` to take
+//!   it away. An echo session meters nothing, so it draws no usage bar until
+//!   this is set; what it exists for is the states that bar can be in, which
+//!   otherwise cost real quota to reach. `/usage 42`, `/usage 95 20`,
+//!   `/usage 42 never`, `/usage notloggedin`, `/usage unreachable`,
+//!   `/usage failed`. The vocabulary is `usage::Fixture`'s, where the states
+//!   live.
+//! - `/limit [minutes]` -- a turn that stops because the account is out of
+//!   quota, saying the limit lifts in `minutes` (default 5, and `never` for a
+//!   limit with no stated reset). What it exists for is auto-resume, which is
+//!   otherwise reachable only by actually exhausting somebody's account: pair
+//!   it with `/usage 100 5` for a meter that agrees, and then `/usage 20` for
+//!   the moment the limit lifts. The wait itself is decided by the meter, so
+//!   those two commands are the whole rig.
 //! - `/compact` -- a compaction, start to finish.
 //! - `/stream N` -- one long answer in N small pieces, 50ms apart: the shape a
 //!   real model's reply arrives in, and the one where the row a reader is
@@ -34,6 +48,10 @@
 //!   and height the app draws, in one session, which is what a scrolling
 //!   problem needs in order to be reproduced twice the same way.
 //! - `/table [columns]` -- a markdown table with cells too long for one line.
+//! - `/subagent [n]` -- n subagents at once (default 1), each named
+//!   "helper k", its prompt recorded as its own first user message: a
+//!   streamed reply, one Bash call, then it finishes about three seconds
+//!   later, the same lifecycle a real Task call has -- see `SUBAGENTS.md`.
 //!
 //! `/slow` earns its place: a queued message, a Stop button and a spinner are
 //! states that only exist mid-turn, and the obvious way to get one -- ask a
@@ -48,6 +66,7 @@ use std::time::Duration;
 use super::driver::{
    AttachmentRef, Driver, Event, EventSink, QuestionOption, SessionStatus, Unqueued,
 };
+use super::subagent::Subagents;

 /// Delay between streamed deltas -- long enough that streaming is visibly
 /// streaming, short enough that tests waiting on a full turn stay fast.
@@ -85,11 +104,19 @@ pub struct EchoDriver {
    /// because `/ask` puts up to four on one tool call, and the turn resumes
    /// when the last is answered rather than the first.
    pending_questions: Mutex<Vec<PendingQuestion>>,
+    /// The invented rate-limit answer `/usage` sets, shared with the usage
+    /// monitor that serves it. An echo session meters nothing, so this is unset
+    /// until a test asks for something -- see [`crate::usage::Fixture`].
+    usage: crate::usage::Fixture,
    /// A pretend context, so the status row has something that behaves the way
    /// a real one does: it grows with each turn, drops to what the compaction
    /// says it recovered, and a clear leaves it unmeasured. What is real is
    /// which way the numbers move.
    context: Arc<AtomicU64>,
+    /// This session's subagents -- see `SUBAGENTS.md`. `/subagent` is the
+    /// test rig for the same registry the claude driver routes real Task
+    /// calls into.
+    subagents: Arc<Subagents>,
 }

 impl EchoDriver {
@@ -311,6 +338,110 @@ impl EchoDriver {
            return;
        }

+        // Answered here rather than in the turn below, because it is not a
+        // turn: nothing is generated, and what is being exercised is the
+        // *other* screens -- the bar under the header, the button beside it and
+        // the dialog it opens, which read the usage route, not this transcript.
+        if let Some(rest) = text.strip_prefix("/usage") {
+            if announce {
+                self.emit(Event::MessageTaken {
+                    id: None,
+                    text: text.clone(),
+                    attachments,
+                });
+            }
+            let said = self.usage.command(rest);
+            self.emit(Event::AssistantText {
+                delta: format!("{said}\n"),
+            });
+            self.emit(Event::Status {
+                state: SessionStatus::Idle,
+            });
+            return;
+        }
+
+        // A turn that ends the way a real one does when the account runs out:
+        // the same event a real driver reports, so what acts on it -- the
+        // transcript row and `crate::resume` -- is exercised rather than
+        // imitated. The meter it should agree with is `/usage`'s fixture,
+        // deliberately separate: the two disagreeing is a state worth being
+        // able to produce, since it is what a stale reset time looks like.
+        if let Some(rest) = text.strip_prefix("/limit") {
+            if announce {
+                self.emit(Event::MessageTaken {
+                    id: None,
+                    text: text.clone(),
+                    attachments,
+                });
+            }
+            let rest = rest.trim();
+            let resets_at = match rest {
+                "never" | "none" => None,
+                "" => Some(super::now() + 5.0 * 60.0),
+                minutes => Some(super::now() + minutes.parse::<f64>().unwrap_or(5.0) * 60.0),
+            };
+            self.emit(Event::Status {
+                state: SessionStatus::Running,
+            });
+            self.emit(Event::AssistantText {
+                delta: "Working on it".to_string(),
+            });
+            self.emit(Event::LimitReached { resets_at });
+            self.emit(Event::Status {
+                state: SessionStatus::Idle,
+            });
+            return;
+        }
+
+        // `n` subagents at once, each with its own transcript in the
+        // registry a real Task call routes into -- see `SUBAGENTS.md`. The
+        // parent's own Task calls end when their subagent does, three
+        // seconds later, which is long enough to see the running state on
+        // the phone before it finishes.
+        if let Some(rest) = text.strip_prefix("/subagent") {
+            let n = rest.trim().parse::<usize>().unwrap_or(1).clamp(1, 8);
+            if announce {
+                self.emit(Event::MessageTaken {
+                    id: None,
+                    text: text.clone(),
+                    attachments,
+                });
+            }
+            self.emit(Event::Status {
+                state: SessionStatus::Running,
+            });
+            let sink = self.sink.clone();
+            let subagents = Arc::clone(&self.subagents);
+            tokio::spawn(async move {
+                let mut helpers = Vec::new();
+                for k in 1..=n {
+                    let id = format!("echo-subagent-{k}-{}", super::random_hex());
+                    let title = format!("helper {k}");
+                    let prompt = format!(
+                        "You are helper {k} of {n}. Say a few words, run a command, then stop."
+                    );
+                    let _ = sink.send(Event::ToolStart {
+                        id: id.clone(),
+                        tool: "Task".to_string(),
+                        input: serde_json::json!({
+                            "description": title,
+                            "prompt": prompt,
+                            "subagent_type": "general-purpose",
+                        }),
+                    });
+                    subagents.start(&id, &title, Some(&prompt));
+                    helpers.push((id, sink.clone(), Arc::clone(&subagents)));
+                }
+                for (id, sink, subagents) in helpers {
+                    tokio::spawn(run_helper(id, sink, subagents));
+                }
+                let _ = sink.send(Event::Status {
+                    state: SessionStatus::Idle,
+                });
+            });
+            return;
+        }
+
        // The same word the real CLI takes, so a phone drives both the same way.
        // `Driver::compact` is what the manager's route calls; this is the typed
        // path onto it.
@@ -606,7 +737,12 @@ impl EchoDriver {
        });
    }

-    pub fn new(sink: EventSink, session_dir: PathBuf) -> Self {
+    pub fn new(
+        sink: EventSink,
+        session_dir: PathBuf,
+        usage: crate::usage::Fixture,
+        subagents: Arc<Subagents>,
+    ) -> Self {
        let driver = Self {
            sink,
            pending_questions: Mutex::new(Vec::new()),
@@ -614,6 +750,8 @@ impl EchoDriver {
            busy: Arc::new(AtomicBool::new(false)),
            queued: Arc::new(Mutex::new(Vec::new())),
            session_dir,
+            usage,
+            subagents,
        };
        driver.emit(Event::Status {
            state: SessionStatus::Idle,
@@ -735,6 +873,51 @@ async fn write_beat(sink: &EventSink, session_dir: &Path, beat: usize) {
    tokio::time::sleep(Duration::from_millis(120)).await;
 }

+/// One `/subagent` helper: a few streamed words, one Bash call, then
+/// `Status::Exited` about three seconds after it started -- long enough that
+/// its `Running` state can be seen on the phone before it finishes. The
+/// parent's own Task call for it ends at the same moment, the same way a
+/// real Task's `tool_result` ends it.
+async fn run_helper(id: String, sink: EventSink, subagents: Arc<Subagents>) {
+    let start = tokio::time::Instant::now();
+    for word in "Working on it now.".split_inclusive(' ') {
+        subagents.record(
+            &id,
+            Event::AssistantText {
+                delta: word.to_string(),
+            },
+        );
+        tokio::time::sleep(DELTA_DELAY).await;
+    }
+    let tool_id = format!("{id}-bash");
+    subagents.record(
+        &id,
+        Event::ToolStart {
+            id: tool_id.clone(),
+            tool: "Bash".to_string(),
+            input: serde_json::json!({ "command": "echo helper done" }),
+        },
+    );
+    tokio::time::sleep(DELTA_DELAY).await;
+    subagents.record(
+        &id,
+        Event::ToolEnd {
+            id: tool_id,
+            output: "helper done".to_string(),
+        },
+    );
+    let target = Duration::from_secs(3);
+    let elapsed = start.elapsed();
+    if elapsed < target {
+        tokio::time::sleep(target - elapsed).await;
+    }
+    subagents.finish(&id);
+    let _ = sink.send(Event::ToolEnd {
+        id,
+        output: "subagent finished".to_string(),
+    });
+}
+
 /// A message written during a turn and waiting for it to end: the id of the
 /// `MessageQueued` that announced it, what it said, and what was attached. All
 /// three, because all three are what the `MessageTaken` at the other end owes.
@@ -5,9 +5,19 @@
 //!
 //! **It is spawned but not spoken to over stdio.** The process is started
 //! through the same [`Transport`] as any other and then reached over HTTP on a
-//! loopback port. A remote llama-server would need its port forwarded as well
-//! as its command wrapped, which is not built, so a session on an ssh host is
-//! refused rather than silently talking to the wrong machine.
+//! loopback port. That is the second half of what a transport is -- "run this"
+//! plus "reach this port" -- and it is what lets a session run on another
+//! machine: [`Transport::reserve_port`] hands back a port the server binds
+//! *there* and one that reaches it *here*, and the ssh connection carrying the
+//! command carries the tunnel between them. The far `llama-server` binds
+//! loopback only, so a model is never served to that machine's network.
+//!
+//! **The model file is the far machine's, not this one's.** A remote setup
+//! names its own models directory (`SshConfig::models_dir`, defaulting to where
+//! this backend keeps its downloads), and the file is looked for *there* -- so
+//! a session naming a model that machine does not have says so, instead of
+//! starting a server that will never load one. Downloading to another machine
+//! is not built; the model gets there however anything else does.
 //!
 //! **The server is stateless between requests**, so the whole conversation goes
 //! with every one. It is rebuilt from the session's transcript rather than kept
@@ -67,6 +77,7 @@ impl LlamaDriver {
    /// in a different currency: two servers holding the same model is twice the
    /// memory, and the second would bind a different port while the phone kept
    /// talking to the first.
+    #[allow(clippy::too_many_arguments)]
    pub fn launch(
        meta: &SessionConfig,
        provider: &ProviderConfig,
@@ -75,17 +86,15 @@ impl LlamaDriver {
        transcript: &Path,
        session_dir: &Path,
        sink: EventSink,
+        // llama.cpp has no notion of a Task call, so this is accepted only
+        // to keep one shape across every driver's launch -- see
+        // `SUBAGENTS.md`'s "Server layout".
+        _subagents: Arc<super::subagent::Subagents>,
    ) -> Result<Self> {
-        if !matches!(transport, Transport::Here) {
-            bail!(
-                "llama.cpp sessions can only run on this machine for now: the model is served \
-                 over HTTP, and forwarding that port to another host isn't built yet."
-            );
-        }
        let model = meta.model.as_deref().context(
            "a llama.cpp session needs a model -- one of the downloaded ones, by its key",
        )?;
-        let path = model_path(models_dir, model)?;
+        let path = model_on(transport, models_dir, model)?;

        // Already loaded and still running: keep talking to it. The health poll
        // below confirms it is really answering, so adopting a pid whose server
@@ -111,14 +120,21 @@ impl LlamaDriver {
            ));
        }

-        let port = free_port().context("finding a port for llama-server")?;
+        // Where it listens on its own machine, and where that is reached
+        // from here -- the same number when that machine is this one.
+        let forward = transport
+            .reserve_port()
+            .context("finding a port for llama-server")?;
        let mut args: Vec<String> = vec![
            "-m".into(),
-            path.to_string_lossy().into_owned(),
+            path.clone(),
+            // Loopback there, whichever machine there is: what reaches it
+            // from outside that machine is the ssh tunnel and nothing
+            // else.
            "--host".into(),
            "127.0.0.1".into(),
            "--port".into(),
-            port.to_string(),
+            forward.there.to_string(),
        ];
        // Settings that belong to the server because they decide how the model
        // is loaded; the sampling ones ride on each request instead, so changing
@@ -134,8 +150,8 @@ impl LlamaDriver {
            }
        }

-        let program = provider.command.as_deref().unwrap_or("llama-server");
-        let launch = Launch::new(program, args, meta.cwd.as_deref());
+        let program = provider.program();
+        let launch = Launch::new(program, args, meta.cwd.as_deref()).reaching(forward);
        // Its output goes to files, not pipes. Not only so the process can
        // outlive this server: nothing ever read those pipes, so a chatty
        // llama-server filled the 64 KB buffer and blocked mid-load with no sign
@@ -152,8 +168,12 @@ impl LlamaDriver {
            .id()
            .context("llama-server exited before it could be recorded")?;
        tracing::info!(
-            "session {} running {program} for {model} on 127.0.0.1:{port} as pid {pid}",
-            meta.id
+            "session {} running {program} for {model} {} on 127.0.0.1:{} there, \
+             reached at 127.0.0.1:{} here, as pid {pid}",
+            meta.id,
+            transport.describe(),
+            forward.there,
+            forward.here,
        );
        // Reaped so it does not become a zombie while this server is still its
        // parent; the health poll and the record are what say whether the
@@ -163,12 +183,18 @@ impl LlamaDriver {
            let _ = child.wait().await;
        });

-        let record = process::Record::of(pid, process::Detail::Http { port })
+        // The *near* port, because that is the one anything reaching this
+        // server has to dial -- including a later run of this backend,
+        // which adopts the record without knowing which machine the server
+        // is on. For a remote session the recorded pid is the ssh
+        // client's, which is the process this machine owns and which holds
+        // the tunnel open for exactly as long as the far server lives.
+        let record = process::Record::of(pid, process::Detail::Http { port: forward.here })
            .context("llama-server was gone before its start time could be read")?;
        process::write(session_dir, &record);

        Ok(Self::attached(
-            format!("http://127.0.0.1:{port}"),
+            format!("http://127.0.0.1:{}", forward.here),
            meta,
            model,
            transcript,
@@ -202,7 +228,7 @@ impl LlamaDriver {
            let endpoint = endpoint.clone();
            let model = model.to_string();
            let session_dir = session_dir.to_path_buf();
-            std::thread::spawn(move || match wait_until_ready(&endpoint) {
+            std::thread::spawn(move || match wait_until_ready(&endpoint, &session_dir) {
                Ok(()) => {
                    tracing::info!("{model} loaded and answering at {endpoint}");
                    let _ = sink.send(Event::Status {
@@ -499,17 +525,70 @@ fn model_path(models_dir: &Path, key: &str) -> Result<PathBuf> {
    Ok(path)
 }

-/// An unused loopback port, by asking the OS for one and letting it go. Racy in
-/// principle, but nothing on this machine is hunting for ports, and the
-/// alternative -- parsing the port back out of the server's log -- couples us to
-/// its output format for no real gain.
-fn free_port() -> Result<u16> {
-    let listener = std::net::TcpListener::bind("127.0.0.1:0")?;
-    Ok(listener.local_addr()?.port())
+/// The model file's path **on the machine that will serve it**, confirmed to be
+/// there.
+///
+/// One function rather than a local check and hope for the other case: the same
+/// question has to be asked of two filesystems. The remote answer is measured
+/// for the reason the local one is -- a missing file otherwise becomes a
+/// `llama-server` that starts, fails to load, and reports as a session that
+/// never became ready, which reads as the machine being slow.
+///
+/// One blocking round trip on a remote spawn, which is what the spawn is
+/// already paying to start ssh. The alternative is a path built here from a `~`
+/// this machine cannot expand.
+fn model_on(transport: &Transport, models_dir: &Path, key: &str) -> Result<String> {
+    let Transport::Ssh { name, .. } = transport else {
+        return Ok(model_path(models_dir, key)?.to_string_lossy().into_owned());
+    };
+    // The same directory the spawn screen listed for this machine, and one
+    // function for the same reason: a list from one place and a load from
+    // another is a model that appears and then fails.
+    let dir = crate::models::dir_on(transport, models_dir);
+    // Checked here rather than in the script: `..` in a key would walk out of
+    // the models directory on a machine this server can start processes on,
+    // and the phone is where the key comes from.
+    for part in key.split('/') {
+        if part.is_empty() || part == "." || part == ".." {
+            bail!("\"{key}\" is not a model key this can resolve");
+        }
+    }
+    let path = format!("{}/{key}", dir.trim_end_matches('/'));
+    // `$HOME` on the far side, which is the only machine that knows what it is,
+    // and the resolved path printed back so the launch hands `llama-server`
+    // something absolute. "Not there" is answered rather than failed, because a
+    // machine that could not be asked at all has to say so in its own words --
+    // it would otherwise arrive as this same sentence about a missing model.
+    let script = "p=$1; case $p in \"~\") p=$HOME;; \"~/\"*) p=$HOME/${p#\"~/\"};; esac; \
+                  [ -f \"$p\" ] && printf 'at\\t%s\\n' \"$p\" || printf 'missing\\n'"
+        .to_string();
+    let launch = Launch::new(
+        "sh",
+        vec!["-c".to_string(), script, "sh".to_string(), path.clone()],
+        None,
+    );
+    let answer = transport
+        .capture_blocking(&launch)
+        .with_context(|| format!("couldn't ask {name} where its models are"))?;
+    match answer.trim().split_once('\t') {
+        Some(("at", resolved)) => Ok(resolved.to_string()),
+        _ => bail!(
+            "{name} has no model at {path}. A llama.cpp session serves the file from the \
+             machine it runs on, so the model has to be on {name} -- what this backend has \
+             downloaded is somewhere else."
+        ),
+    }
 }

 /// Polls until the server says it is ready, or gives up.
-fn wait_until_ready(endpoint: &str) -> Result<()> {
+///
+/// Watches the process as well as the port, because the two failures need
+/// different words and one of them is common: a model that will not load,
+/// a port already taken on the far machine, a `llama-server` too old for
+/// a flag. All of those exit within a second and none of them will ever
+/// answer `/health`, so waiting out the timeout turns a server that said
+/// exactly what was wrong into "gave up after 300s".
+fn wait_until_ready(endpoint: &str, session_dir: &Path) -> Result<()> {
    let deadline = std::time::Instant::now() + READY_TIMEOUT;
    let url = format!("{endpoint}/health");
    loop {
@@ -518,13 +597,48 @@ fn wait_until_ready(endpoint: &str) -> Result<()> {
        {
            return Ok(());
        }
+        // `None` is the session having been stopped or deleted while this
+        // waited, which is nobody's fault and still not worth waiting on.
+        match process::recorded(session_dir) {
+            Some((_, process::Liveness::Alive | process::Liveness::Unknown)) => {}
+            Some((_, process::Liveness::Dead)) | None => {
+                bail!("it exited before it answered.{}", log_tail(session_dir));
+            }
+        }
        if std::time::Instant::now() > deadline {
-            bail!("gave up after {}s", READY_TIMEOUT.as_secs());
+            bail!(
+                "gave up after {}s.{}",
+                READY_TIMEOUT.as_secs(),
+                log_tail(session_dir)
+            );
        }
        std::thread::sleep(std::time::Duration::from_millis(250));
    }
 }

+/// The end of `llama-server`'s own log, for a failure message.
+///
+/// Its account of what went wrong is the useful half -- "failed to load
+/// model", "bind: Address already in use" -- and on a remote session it
+/// is the only half, since nobody reading the phone can open a file on
+/// that machine. Bounded, because this ends up in an event a phone draws.
+fn log_tail(session_dir: &Path) -> String {
+    let Ok(text) = std::fs::read_to_string(session_dir.join(SERVER_LOG)) else {
+        return String::new();
+    };
+    let tail: Vec<&str> = text.lines().rev().take(LOG_TAIL_LINES).collect();
+    if tail.is_empty() {
+        return String::new();
+    }
+    format!(
+        " It last said: {}",
+        tail.into_iter().rev().collect::<Vec<_>>().join(" / ")
+    )
+}
+
+/// How much of that log to carry into a message somebody reads on a phone.
+const LOG_TAIL_LINES: usize = 6;
+
 /// One streamed completion: posts the conversation, emits each delta as it
 /// arrives. Emits rather than returns, because the transcript those events land
 /// in is what the next turn reads back.
@@ -0,0 +1,507 @@
+//! A session's subagents -- see `SUBAGENTS.md`.
+//!
+//! **A subagent is a second transcript owned by a session, in the same event
+//! model, with no process and no controls.** It shares the transcript file
+//! format, the paging routes, and the SSE stream with a session by
+//! addressing, not by copying: `Transcript`, `read_window` and `catch_up`
+//! work on a subagent's file unchanged.
+//!
+//! Storage is `<session dir>/subagents/<id>/{meta.json,transcript.jsonl}`,
+//! where `<id>` is the Task tool_use id that started it -- unique, stable
+//! across a backend restart, and already the key the parent side uses. Only
+//! ids matching [`is_subagent_id`] are ever turned into a path.
+
+use std::collections::HashMap;
+use std::fs;
+use std::path::{Path, PathBuf};
+use std::sync::{Arc, Mutex};
+
+use anyhow::{Context, Result};
+use serde::{Deserialize, Serialize};
+use tokio::sync::broadcast;
+
+use super::driver::{Event, SessionStatus};
+use super::transcript::{SeqEvent, Transcript};
+
+/// Fan-out buffer for one subagent's SSE subscribers. Smaller than a
+/// session's: a subagent's whole conversation is usually a handful of tool
+/// calls, not an hours-long session.
+const EVENT_BUFFER: usize = 64;
+
+/// Whether `id` is safe to become a path segment under a session's
+/// `subagents/` directory. Mirrors `import::is_session_id`'s reasoning: the
+/// id arrives as a value inside JSON the CLI sent, and it becomes a
+/// directory name, so a `/` or `..` in it must never be trusted.
+fn is_subagent_id(id: &str) -> bool {
+    !id.is_empty()
+        && id.len() <= 200
+        && id
+            .bytes()
+            .all(|b| b.is_ascii_alphanumeric() || b == b'_' || b == b'-')
+}
+
+/// What a subagent's directory holds beside its transcript. Small and
+/// separate from `Subagent` itself because this is exactly what survives a
+/// backend restart on disk, and nothing else does.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+struct Meta {
+    title: String,
+    /// Epoch seconds. Absent from `SubagentInfo`'s sort key deliberately:
+    /// `list` sorts by this rather than by directory order, which a
+    /// filesystem does not promise.
+    created: f64,
+}
+
+/// One row of `GET /sessions/{id}/subagents`.
+#[derive(Debug, Clone, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct SubagentInfo {
+    pub id: String,
+    pub title: String,
+    pub status: SessionStatus,
+    pub created: f64,
+    pub last_activity: f64,
+}
+
+/// One subagent: its own transcript and broadcast, same shape as a
+/// session's but with no driver behind it.
+pub struct Subagent {
+    dir: PathBuf,
+    transcript: Mutex<Transcript>,
+    events: broadcast::Sender<SeqEvent>,
+    /// Mirrors the transcript's last `Status` event, kept live rather than
+    /// read back from `Transcript::last_status` -- that answers "as of
+    /// opening" (see its own doc comment) and never moves for an append made
+    /// through *this* object, which is every append a live subagent ever
+    /// makes. Without this, `finish` immediately after `start` in the same
+    /// process read the file's stale opening status and reported itself
+    /// still open.
+    status: Mutex<SessionStatus>,
+}
+
+impl Subagent {
+    pub fn transcript_path(&self) -> PathBuf {
+        self.dir.join("transcript.jsonl")
+    }
+
+    pub fn subscribe(&self) -> broadcast::Receiver<SeqEvent> {
+        self.events.subscribe()
+    }
+
+    /// Whether this subagent's last recorded status is not `Exited` --
+    /// what decides whether a further child line reopens it (see
+    /// `Subagents::reopen`) rather than continuing straight through. See
+    /// `SUBAGENTS.md`'s lifecycle.
+    pub fn is_open(&self) -> bool {
+        *self.status.lock().unwrap() != SessionStatus::Exited
+    }
+
+    fn append(&self, event: Event) {
+        let mut transcript = self.transcript.lock().unwrap();
+        match transcript.append(event, super::now()) {
+            Ok(entry) => {
+                if let Event::Status { state } = &entry.event {
+                    *self.status.lock().unwrap() = *state;
+                }
+                // No subscribers is fine; the transcript already has it,
+                // same as a session's pump.
+                let _ = self.events.send(entry);
+            }
+            Err(err) => tracing::error!("subagent transcript append failed: {err:#}"),
+        }
+    }
+}
+
+/// Every subagent one session has started, keyed by the Task tool_use id
+/// that names it.
+///
+/// Lives beside a session's driver rather than inside it: a claude driver
+/// holds an `Arc` to this and routes child lines into it; echo uses it for
+/// its `/subagent` rig; llama ignores it, since it has no notion of a Task
+/// call. One instance per live session, built at launch and handed to
+/// whichever driver replaces it across a stop/start.
+pub struct Subagents {
+    /// The session's own directory; subagents live under `<dir>/subagents`.
+    dir: PathBuf,
+    live: Mutex<HashMap<String, Arc<Subagent>>>,
+}
+
+impl Subagents {
+    pub fn new(session_dir: PathBuf) -> Self {
+        Self {
+            dir: session_dir,
+            live: Mutex::new(HashMap::new()),
+        }
+    }
+
+    fn subagents_dir(&self) -> PathBuf {
+        self.dir.join("subagents")
+    }
+
+    /// Opens the subagent named `id`, creating it if this is the first
+    /// anyone has heard of it -- on disk as well as in memory, so a
+    /// subagent from before a backend restart is reopened rather than
+    /// recreated. `title`/`prompt` are used only at creation: reopening an
+    /// existing one keeps its original title and never repeats the prompt
+    /// into its transcript a second time.
+    fn open_or_create(&self, id: &str, title: &str, prompt: Option<&str>) -> Result<Arc<Subagent>> {
+        let dir = self.subagents_dir().join(id);
+        let meta_path = dir.join("meta.json");
+        let existed = meta_path.is_file();
+        let meta = if existed {
+            let text = fs::read_to_string(&meta_path)
+                .with_context(|| format!("read {}", meta_path.display()))?;
+            serde_json::from_str::<Meta>(&text).context("parse subagent meta")?
+        } else {
+            wg_app_link::private::create_dir(&dir)?;
+            let meta = Meta {
+                title: title.to_string(),
+                created: super::now(),
+            };
+            wg_app_link::private::write_file(
+                &meta_path,
+                serde_json::to_string(&meta)
+                    .context("serialize subagent meta")?
+                    .as_bytes(),
+            )?;
+            meta
+        };
+        let mut transcript = Transcript::open(&dir.join("transcript.jsonl"))?;
+        if !existed {
+            // First lines, in order: the subagent is running the moment it
+            // exists, and its prompt -- when known -- is genuinely its first
+            // user turn. Written once, here, so a reopen never repeats them.
+            transcript.append(
+                Event::Status {
+                    state: SessionStatus::Running,
+                },
+                meta.created,
+            )?;
+            if let Some(prompt) = prompt {
+                transcript.append(
+                    Event::UserMessage {
+                        id: None,
+                        text: prompt.to_string(),
+                        attachments: Vec::new(),
+                    },
+                    meta.created,
+                )?;
+            }
+        }
+        // A freshly created subagent is running by construction (its only
+        // lines so far are `Status::Running` and maybe its prompt); a
+        // reopened one takes whatever the file last said, since this
+        // `Transcript` has not been appended to yet in this process.
+        let status = if existed {
+            transcript.last_status().unwrap_or(SessionStatus::Running)
+        } else {
+            SessionStatus::Running
+        };
+        let (events, _) = broadcast::channel(EVENT_BUFFER);
+        Ok(Arc::new(Subagent {
+            dir,
+            transcript: Mutex::new(transcript),
+            events,
+            status: Mutex::new(status),
+        }))
+    }
+
+    /// Starts a subagent unless one is already known by this id -- see
+    /// `SUBAGENTS.md`'s lifecycle: created at the Task call or at the first
+    /// child line, whichever comes first, and never twice. A bad id is
+    /// refused rather than turned into a path.
+    pub fn start(&self, id: &str, title: &str, prompt: Option<&str>) {
+        if !is_subagent_id(id) {
+            tracing::debug!("refusing to start a subagent with a bad id {id:?}");
+            return;
+        }
+        let mut live = self.live.lock().unwrap();
+        if live.contains_key(id) {
+            return;
+        }
+        match self.open_or_create(id, title, prompt) {
+            Ok(subagent) => {
+                live.insert(id.to_string(), subagent);
+            }
+            Err(err) => tracing::error!("couldn't start subagent {id}: {err:#}"),
+        }
+    }
+
+    /// The subagent named `id`, reopening it from disk on first use in this
+    /// process if one is there. `None` for an id nothing has ever started --
+    /// deliberately not created here, since a route or a routing decision is
+    /// not the Task call that is supposed to be the only way one begins.
+    pub fn get(&self, id: &str) -> Option<Arc<Subagent>> {
+        if !is_subagent_id(id) {
+            return None;
+        }
+        if let Some(existing) = self.live.lock().unwrap().get(id).cloned() {
+            return Some(existing);
+        }
+        if !self.subagents_dir().join(id).join("meta.json").is_file() {
+            return None;
+        }
+        // Title and prompt are ignored: the directory already exists, so
+        // `open_or_create` reads its own meta rather than using either.
+        match self.open_or_create(id, "", None) {
+            Ok(subagent) => {
+                self.live
+                    .lock()
+                    .unwrap()
+                    .insert(id.to_string(), Arc::clone(&subagent));
+                Some(subagent)
+            }
+            Err(err) => {
+                tracing::error!("couldn't reopen subagent {id}: {err:#}");
+                None
+            }
+        }
+    }
+
+    /// Appends one event to a subagent's own transcript. A no-op, with a
+    /// debug log, for an id nothing was started under -- a child line for a
+    /// subagent this registry never opened is dropped rather than guessed
+    /// at.
+    pub fn record(&self, id: &str, event: Event) {
+        match self.live.lock().unwrap().get(id).cloned() {
+            Some(subagent) => subagent.append(event),
+            None => tracing::debug!("dropping an event for unknown subagent {id}"),
+        }
+    }
+
+    /// The subagent's own turn ended: its `Status::Exited`. Called from
+    /// `translate_child` on the subagent's own `end_turn`, never on the
+    /// parent's `tool_result` -- a background Task's `tool_result` arrives
+    /// at launch, not at completion, so it says nothing about whether this
+    /// is over. A no-op for an id that is not a subagent's or is already
+    /// closed.
+    pub fn finish(&self, id: &str) {
+        if let Some(subagent) = self.live.lock().unwrap().get(id).cloned()
+            && subagent.is_open()
+        {
+            subagent.append(Event::Status {
+                state: SessionStatus::Exited,
+            });
+        }
+    }
+
+    /// A line arrived for a subagent that had already finished: it is
+    /// working again, not stale -- a background Task can be sent another
+    /// message long after its first turn ended. Appends `Status::Running`
+    /// so the list stops reporting it as finished; a no-op if it was not
+    /// actually closed, so a caller need not check first.
+    pub fn reopen(&self, id: &str) {
+        if let Some(subagent) = self.live.lock().unwrap().get(id).cloned()
+            && !subagent.is_open()
+        {
+            subagent.append(Event::Status {
+                state: SessionStatus::Running,
+            });
+        }
+    }
+
+    /// The parent session's process is gone, so nothing still open here has
+    /// a process behind it either -- see `SUBAGENTS.md`'s lifecycle #4.
+    pub fn finish_all(&self) {
+        let subagents: Vec<Arc<Subagent>> = self.live.lock().unwrap().values().cloned().collect();
+        for subagent in subagents {
+            if subagent.is_open() {
+                subagent.append(Event::Status {
+                    state: SessionStatus::Exited,
+                });
+            }
+        }
+    }
+
+    /// Every subagent under this session's directory, oldest first --
+    /// `GET /sessions/{id}/subagents`. Read straight from disk rather than
+    /// from `live`, so a subagent from before this process started (or one
+    /// this run has not yet touched) still shows up; one file read per
+    /// subagent, which is fine at the handful a session usually has.
+    ///
+    /// `session_running` is what turns a subagent whose last status is
+    /// `Running` into `Unknown`: its process was the session's, and the
+    /// session has none.
+    pub fn list(&self, session_running: bool) -> Vec<SubagentInfo> {
+        let mut rows: Vec<SubagentInfo> = match fs::read_dir(self.subagents_dir()) {
+            Ok(entries) => entries
+                .filter_map(Result::ok)
+                .filter_map(|entry| info_of(&entry.path(), session_running))
+                .collect(),
+            // No directory is no subagents, not a fault worth reporting.
+            Err(_) => Vec::new(),
+        };
+        rows.sort_by(|a, b| {
+            a.created
+                .partial_cmp(&b.created)
+                .unwrap_or(std::cmp::Ordering::Equal)
+        });
+        rows
+    }
+}
+
+fn info_of(subagent_dir: &Path, session_running: bool) -> Option<SubagentInfo> {
+    let id = subagent_dir.file_name()?.to_str()?.to_string();
+    let meta_path = subagent_dir.join("meta.json");
+    let text = fs::read_to_string(&meta_path).ok()?;
+    let meta: Meta = serde_json::from_str(&text).ok()?;
+    let transcript = Transcript::open(&subagent_dir.join("transcript.jsonl")).ok()?;
+    // The subagent's first line is always `Status::Running`, written before
+    // this directory is discoverable at all, so `None` here is not a state a
+    // reader can actually observe -- but it is not this function's place to
+    // invent one, so a status this build does not expect to see falls back
+    // to the word the lifecycle promises it started in.
+    let last_status = transcript.last_status().unwrap_or(SessionStatus::Running);
+    let status = if last_status == SessionStatus::Running && !session_running {
+        SessionStatus::Unknown
+    } else {
+        last_status
+    };
+    Some(SubagentInfo {
+        id,
+        title: meta.title,
+        status,
+        created: meta.created,
+        last_activity: transcript.last_activity().unwrap_or(meta.created),
+    })
+}
+
+/// How many subagents a session has, for `SessionInfo::subagents`: a
+/// directory listing, so the session list stays cheap and only the
+/// dedicated route pays for reading a status out of each one.
+pub fn count(session_dir: &Path) -> usize {
+    fs::read_dir(session_dir.join("subagents"))
+        .map(|entries| entries.filter_map(Result::ok).count())
+        .unwrap_or(0)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn a_bad_id_is_refused_rather_than_turned_into_a_path() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let subagents = Subagents::new(dir.path().to_path_buf());
+        subagents.start("../../etc", "escape", None);
+        assert!(subagents.get("../../etc").is_none());
+        assert!(!dir.path().join("subagents").exists());
+    }
+
+    #[test]
+    fn starting_twice_keeps_the_first_title_and_does_not_repeat_the_prompt() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let subagents = Subagents::new(dir.path().to_path_buf());
+        subagents.start("toolu_1", "first title", Some("do the thing"));
+        subagents.start("toolu_1", "second title", Some("do the thing"));
+
+        let rows = subagents.list(true);
+        assert_eq!(rows.len(), 1);
+        assert_eq!(rows[0].title, "first title");
+
+        let events = crate::session::transcript::read_after(
+            &subagents.get("toolu_1").unwrap().transcript_path(),
+            0,
+        )
+        .expect("read");
+        assert_eq!(
+            events
+                .iter()
+                .filter(|e| matches!(e.event, Event::UserMessage { .. }))
+                .count(),
+            1
+        );
+    }
+
+    #[test]
+    fn a_reopened_subagent_continues_its_own_transcript() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        {
+            let subagents = Subagents::new(dir.path().to_path_buf());
+            subagents.start("toolu_2", "helper", Some("go"));
+            subagents.record(
+                "toolu_2",
+                Event::AssistantText {
+                    delta: "working".to_string(),
+                },
+            );
+        }
+        // A fresh registry, the way a backend restart builds one.
+        let subagents = Subagents::new(dir.path().to_path_buf());
+        let subagent = subagents.get("toolu_2").expect("reopened");
+        assert!(subagent.is_open());
+        subagents.record(
+            "toolu_2",
+            Event::AssistantText {
+                delta: " more".to_string(),
+            },
+        );
+        let events =
+            crate::session::transcript::read_after(&subagent.transcript_path(), 0).expect("read");
+        // Status, UserMessage, two AssistantText deltas, seq continuing.
+        assert_eq!(events.len(), 4);
+        assert_eq!(events.last().unwrap().seq, 4);
+    }
+
+    #[test]
+    fn finishing_appends_exited_and_further_lines_are_droppable() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let subagents = Subagents::new(dir.path().to_path_buf());
+        subagents.start("toolu_3", "helper", None);
+        subagents.finish("toolu_3");
+        let subagent = subagents.get("toolu_3").unwrap();
+        assert!(!subagent.is_open());
+        // On disk too, not only in the live cache `is_open` reads.
+        assert_eq!(
+            Transcript::open(&subagent.transcript_path())
+                .expect("reopen")
+                .last_status(),
+            Some(SessionStatus::Exited)
+        );
+
+        // Finishing an id that was never a subagent is a no-op, not a panic.
+        subagents.finish("never-started");
+    }
+
+    #[test]
+    fn finish_all_closes_only_what_is_still_open() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let subagents = Subagents::new(dir.path().to_path_buf());
+        subagents.start("toolu_4", "one", None);
+        subagents.start("toolu_5", "two", None);
+        subagents.finish("toolu_4");
+        subagents.finish_all();
+
+        let rows = subagents.list(false);
+        assert_eq!(rows.len(), 2);
+        for row in rows {
+            assert_eq!(row.status, SessionStatus::Exited);
+        }
+    }
+
+    #[test]
+    fn a_subagent_still_running_when_the_session_is_not_reports_unknown() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let subagents = Subagents::new(dir.path().to_path_buf());
+        subagents.start("toolu_6", "helper", None);
+
+        assert_eq!(subagents.list(true)[0].status, SessionStatus::Running);
+        assert_eq!(subagents.list(false)[0].status, SessionStatus::Unknown);
+    }
+
+    #[test]
+    fn list_is_oldest_first_and_the_count_matches_the_directory() {
+        let dir = tempfile::tempdir().expect("tempdir");
+        let subagents = Subagents::new(dir.path().to_path_buf());
+        assert_eq!(count(dir.path()), 0);
+        subagents.start("toolu_a", "a", None);
+        std::thread::sleep(std::time::Duration::from_millis(2));
+        subagents.start("toolu_b", "b", None);
+        let rows = subagents.list(true);
+        assert_eq!(
+            rows.iter().map(|r| r.id.as_str()).collect::<Vec<_>>(),
+            ["toolu_a", "toolu_b"]
+        );
+        assert_eq!(count(dir.path()), 2);
+    }
+}
@@ -11,9 +11,13 @@
 //! `crate::ssh`'s: this module decides *which* transport, that one knows what a
 //! correct ssh invocation is.
 //!
-//! Known second operation, not built because nothing needs it yet: a managed
-//! `llama-server` is spawned as a process but then spoken to over HTTP, so a
-//! remote one needs a forwarded port (`ssh -L`) as well.
+//! A transport is therefore two operations rather than one: **run this** and
+//! **reach this port**. The second is what a managed `llama-server` needs -- it
+//! is spawned as a process and then spoken to over HTTP -- and it is a no-op
+//! locally, where the port a program binds is already one this machine can
+//! dial. Over ssh it is an `-L` tunnel on the same connection that runs the
+//! command, so the model server binds loopback on the far machine and is never
+//! exposed to its network. See [`Transport::reserve_port`].

 use std::path::{Path, PathBuf};
 use std::process::Stdio;
@@ -22,14 +26,26 @@ use anyhow::{Context, Result};
 use tokio::process::Child;

 use crate::config::SshConfig;
+pub use crate::ssh::Forward;

 /// What a driver needs run in order to exist as a process. Deliberately just
-/// the three things every transport can carry; anything a particular machine
-/// needs is the transport's own configuration, not something a driver states.
+/// what every transport can carry -- the command, where it runs, and a port the
+/// caller needs to reach; anything a particular machine needs is the
+/// transport's own configuration, not something a driver states.
 pub struct Launch {
    pub program: String,
    pub args: Vec<String>,
    pub cwd: Option<PathBuf>,
+    /// A port this program will listen on, and the port that reaches it
+    /// from here -- see [`Transport::reserve_port`], which is the only
+    /// thing that should produce one.
+    ///
+    /// On the launch rather than in [`Transport::spawn`]'s signature
+    /// because it is part of what is being run: a caller that needs to
+    /// reach the process it is starting says so once, where it says
+    /// everything else about it, and every transport reads it the same
+    /// way.
+    pub forward: Option<Forward>,
 }

 impl Launch {
@@ -38,8 +54,16 @@ impl Launch {
            program: program.into(),
            args,
            cwd: cwd.map(Path::to_path_buf),
+            forward: None,
        }
    }
+
+    /// Says that this program serves `forward.there`, and that the caller
+    /// will reach it at `forward.here`.
+    pub fn reaching(mut self, forward: Forward) -> Self {
+        self.forward = Some(forward);
+        self
+    }
 }

 /// How a launched process's standard streams are connected.
@@ -102,6 +126,7 @@ impl Transport {
            &launch.program,
            &launch.args,
            launch.cwd.as_deref(),
+            launch.forward,
        ));
        match streams {
            Streams::Piped => {
@@ -153,12 +178,15 @@ impl Transport {
            Self::Here => None,
            Self::Ssh { ssh, .. } => Some(ssh),
        };
-        let output =
-            crate::ssh::command(host, &launch.program, &launch.args, launch.cwd.as_deref())
-                .output()
-                .with_context(|| {
-                    format!("couldn't run \"{}\" {}", launch.program, self.describe())
-                })?;
+        let output = crate::ssh::command(
+            host,
+            &launch.program,
+            &launch.args,
+            launch.cwd.as_deref(),
+            launch.forward,
+        )
+        .output()
+        .with_context(|| format!("couldn't run \"{}\" {}", launch.program, self.describe()))?;
        if !output.status.success() {
            let stderr = String::from_utf8_lossy(&output.stderr).trim().to_string();
            anyhow::bail!(if stderr.is_empty() {
@@ -217,6 +245,34 @@ impl Transport {
        })
    }

+    /// Picks a port for a launched program to serve on, and the port that
+    /// reaches it from here.
+    ///
+    /// The "reach this port" half of what a transport is. Locally there is
+    /// one port and the OS chooses it, by binding and letting go -- racy
+    /// in principle, and nothing on this machine is hunting for ports.
+    ///
+    /// Over ssh the near end is chosen the same way and the far end is a
+    /// guess, because there is no portable way to ask a machine for a free
+    /// port that does not race with binding it anyway. It is taken from
+    /// [`FAR_PORTS`], below the range Linux hands out to outgoing
+    /// connections, so a collision means something else deliberately
+    /// listening there. That is not silent: the program fails to bind and
+    /// exits, and `session::llama` reports what its log said rather than
+    /// waiting out its readiness timeout.
+    pub fn reserve_port(&self) -> Result<Forward> {
+        let listener = std::net::TcpListener::bind("127.0.0.1:0")
+            .context("asking this machine for a free port")?;
+        let here = listener.local_addr()?.port();
+        Ok(match self {
+            Self::Here => Forward { there: here, here },
+            Self::Ssh { .. } => Forward {
+                there: rand::random_range(FAR_PORTS),
+                here,
+            },
+        })
+    }
+
    /// How to say where this runs, for a log line a person reads.
    pub fn describe(&self) -> String {
        match self {
@@ -226,6 +282,11 @@ impl Transport {
    }
 }

+/// Where a port on another machine is guessed from: high enough to be out
+/// of the way of services, and below the 32768-60999 Linux hands out to
+/// outgoing connections, which is where a guess would most often collide.
+const FAR_PORTS: std::ops::Range<u16> = 20000..30000;
+
 /// What a command is given on its standard input.
 ///
 /// Three cases rather than an `Option<Stdio>` because they are three genuinely
@@ -25,7 +25,10 @@ use crate::session::transport::{Launch, Transport};
 /// session stores.
 const PROBES: &[(&str, &str, DriverKind)] = &[
    ("claude-cli", "claude", DriverKind::ClaudeCli),
-    ("local-llama", "llama-server", DriverKind::LlamaCpp),
+    // Named for the program rather than for where it runs: it runs
+    // wherever the setup is, and "local" was true only while a llama
+    // session could not be spawned on another machine.
+    ("llama-cpp", "llama-server", DriverKind::LlamaCpp),
 ];

 /// Models offered for a discovered Claude CLI. A shortcut list for the spawn
@@ -14,6 +14,22 @@ use std::process::Command;

 use crate::config::SshConfig;

+/// A port on the machine a command runs on, and the port that reaches it from
+/// the backend.
+///
+/// The second half of what a transport is (PLAN.md's SSH section): "run this"
+/// plus "reach this port". Locally the two numbers are one and nothing is
+/// forwarded; over ssh the connection carries an `-L` tunnel, so a model server
+/// binds loopback on the far machine and is never exposed to its network.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct Forward {
+    /// What the launched program should listen on, on its own machine.
+    pub there: u16,
+    /// What this machine connects to. The same number as `there` when the
+    /// program runs here.
+    pub here: u16,
+}
+
 /// Options forced onto every connection. `BatchMode` makes a missing key fail
 /// immediately with a readable message instead of hanging on a password prompt
 /// nothing can answer; the keepalives turn a silently dropped link into a
@@ -39,6 +55,7 @@ pub fn command(
    program: &str,
    args: &[String],
    cwd: Option<&Path>,
+    forward: Option<Forward>,
 ) -> Command {
    let Some(ssh) = remote else {
        let mut command = Command::new(program);
@@ -55,9 +72,31 @@ pub fn command(
    };

    let mut command = Command::new("ssh");
-    // -T: no pty. This carries JSONL, and a pty would rewrite it (echo, CRLF
-    // translation, ^C handling) into something the parser can't read.
-    command.arg("-T");
+    if let Some(forward) = forward {
+        // A forwarded process is not spoken to over stdio, and that changes how
+        // it is shut down. Everything else here is a CLI reading its stdin, so
+        // killing the ssh client ends it; a `llama-server` never reads its own,
+        // so the same kill left it running on the far machine with the model
+        // loaded -- measured 2026-09-04, an orphan per stopped session. A pty
+        // is what makes sshd hang the far side up. `-tt` because this client
+        // has no terminal to inherit one from. The cost is a log that arrives
+        // through a line discipline, which nothing parses.
+        command.arg("-tt");
+        // Loopback at both ends: the far side binds 127.0.0.1, so what it
+        // serves is reachable only through this connection.
+        command.args([
+            "-L",
+            &format!("127.0.0.1:{}:127.0.0.1:{}", forward.here, forward.there),
+        ]);
+        // Without this a forward that cannot be set up is a warning on stderr
+        // and a session that runs anyway, answering nothing -- which would
+        // arrive as "the model never became ready".
+        command.args(["-o", "ExitOnForwardFailure=yes"]);
+    } else {
+        // -T: no pty. This carries JSONL, and a pty would rewrite it (echo,
+        // CRLF translation, ^C handling) into something the parser can't read.
+        command.arg("-T");
+    }
    for option in SSH_OPTIONS {
        command.args(["-o", option]);
    }
@@ -181,6 +220,7 @@ mod tests {
            port: None,
            identity_file: None,
            options: vec![],
+            models_dir: None,
            attachments_dir: None,
        }
    }
@@ -192,6 +232,7 @@ mod tests {
            "claude",
            &args(["-p", "--verbose"]),
            Some(Path::new("/tmp/x")),
+            None,
        );
        assert_eq!(argv(&command), ["claude", "-p", "--verbose"]);
        assert_eq!(command.get_current_dir(), Some(Path::new("/tmp/x")));
@@ -204,6 +245,7 @@ mod tests {
            port: Some(2222),
            identity_file: Some("/home/me/.ssh/id_ai".into()),
            options: vec!["StrictHostKeyChecking=accept-new".to_string()],
+            models_dir: None,
            attachments_dir: None,
        };
        let rendered = argv(&command(
@@ -211,6 +253,7 @@ mod tests {
            "claude",
            &args(["-p", "--model", "haiku"]),
            Some(Path::new("/home/bob/work")),
+            None,
        ));

        assert_eq!(rendered[0], "ssh");
@@ -231,12 +274,59 @@ mod tests {
    #[test]
    fn a_remote_command_without_a_cwd_just_execs() {
        let ssh = bare_host();
-        let rendered = argv(&command(Some(&ssh), "claude", &args(["-p"]), None));
+        let rendered = argv(&command(Some(&ssh), "claude", &args(["-p"]), None, None));
        assert_eq!(rendered.last().unwrap(), "exec 'claude' '-p'");
        // No -i means no IdentitiesOnly: ~/.ssh/config decides instead.
        assert!(!rendered.contains(&"IdentitiesOnly=yes".to_string()));
    }

+    /// The second half of a transport: the connection that runs the command also
+    /// carries the port that reaches it.
+    ///
+    /// Both ends are pinned to loopback, which is what keeps a model server off
+    /// the far machine's network -- asserted rather than trusted, because
+    /// dropping the addresses is a one-word edit that still works on a machine
+    /// nobody else can reach.
+    #[test]
+    fn a_forwarded_port_rides_the_same_connection_as_the_command() {
+        let ssh = bare_host();
+        let rendered = argv(&command(
+            Some(&ssh),
+            "llama-server",
+            &args(["--port", "24242"]),
+            None,
+            Some(Forward {
+                there: 24242,
+                here: 41000,
+            }),
+        ));
+        let forward = rendered
+            .iter()
+            .position(|arg| arg == "-L")
+            .expect("a forward");
+        assert_eq!(rendered[forward + 1], "127.0.0.1:41000:127.0.0.1:24242");
+        assert!(rendered.contains(&"ExitOnForwardFailure=yes".to_string()));
+        // The half that is easy to lose: without a pty the far process outlives
+        // the connection, because nothing closes a stdin it never reads.
+        assert!(rendered.contains(&"-tt".to_string()));
+        assert!(!rendered.contains(&"-T".to_string()));
+        // Options come before the host, or ssh reads them as part of the
+        // remote command.
+        assert!(forward < rendered.len() - 2);
+        assert_eq!(
+            rendered.last().unwrap(),
+            "exec 'llama-server' '--port' '24242'"
+        );
+
+        // Nothing forwarded is nothing added: every other session is one of
+        // these, and an -L on it would bind a port for no reason.
+        let plain = argv(&command(Some(&ssh), "claude", &args(["-p"]), None, None));
+        assert!(!plain.contains(&"-L".to_string()));
+        // And a session that *is* spoken to over stdio keeps its raw pipe.
+        assert!(plain.contains(&"-T".to_string()));
+        assert!(!plain.contains(&"-tt".to_string()));
+    }
+
    /// The one character quoting must not swallow. A working directory typed as
    /// `~/repos/ai-app` was arriving as the literal directory `~`, and the
    /// remote shell reported it missing -- which reads as the path being wrong
@@ -277,7 +367,13 @@ mod tests {
        assert_eq!(expand_home(Path::new("/tmp/~/x")), Path::new("/tmp/~/x"));
        assert_eq!(expand_home(Path::new("~user/x")), Path::new("~user/x"));

-        let local = command(None, "claude", &args(["-p"]), Some(Path::new("~/work")));
+        let local = command(
+            None,
+            "claude",
+            &args(["-p"]),
+            Some(Path::new("~/work")),
+            None,
+        );
        assert_eq!(local.get_current_dir(), Some(home.join("work").as_path()));
    }

@@ -301,7 +397,7 @@ mod tests {
        // tries to close the quote and start a new command.
        let ssh = bare_host();
        let evil = Path::new("/tmp/'; touch /tmp/pwned; '");
-        let rendered = argv(&command(Some(&ssh), "claude", &[], Some(evil)));
+        let rendered = argv(&command(Some(&ssh), "claude", &[], Some(evil), None));
        let script = rendered.last().unwrap();
        assert_eq!(
            script,
@@ -30,13 +30,13 @@
 //! already has over that machine.

 use std::collections::HashMap;
-use std::sync::Mutex;
+use std::sync::{Arc, Mutex};
 use std::time::{Duration, Instant};

 use serde::Serialize;
 use serde_json::Value;

-use crate::config::{DriverKind, SetupConfig};
+use crate::config::SetupConfig;
 use crate::session::transport::{Launch, Transport};

 const USAGE_URL: &str = "https://api.anthropic.com/api/oauth/usage";
@@ -107,10 +107,30 @@ pub struct UsageSnapshot {
    pub fetched_at: f64,
 }

+/// The name of each meter, said in one place because two lists have to
+/// agree on it: [`UsageSnapshot::provider`], which is what `GET /usage`
+/// labels a row with, and [`crate::config::DriverKind::usage_provider`],
+/// which is how a session says which of those rows is about it.
+pub const CLAUDE: &str = "claude";
+/// The invented one, for testing the screens that draw these -- see
+/// [`Fixture`].
+pub const ECHO: &str = "echo";
+
 pub trait UsageProvider: Send + Sync {
    fn name(&self) -> &'static str;
    /// Blocking -- call off the async workers.
    fn fetch(&self) -> UsageSnapshot;
+    /// How long an answer from this one may be reused.
+    ///
+    /// A property of the provider rather than of the cache, because what
+    /// sets it is what asking costs: [`ClaudeUsage`] makes a network call
+    /// against an endpoint that rate-limits impatient callers, and the
+    /// fixture below reads a mutex. Caching the fixture for three minutes
+    /// would mean a test setting a number and watching the old one for
+    /// most of that, which reads exactly like the command not working.
+    fn poll_interval(&self) -> Duration {
+        MIN_POLL_INTERVAL
+    }
 }

 /// Reads the numbers behind Claude Code's `/usage` from one machine, using the
@@ -121,6 +141,10 @@ pub struct ClaudeUsage {
    pub setup_name: String,
    /// How to reach that machine. `Here` for the backend's own.
    pub transport: Transport,
+    /// The CLI to run there, for the one thing this asks of it: refreshing its
+    /// own expired token. The provider's, so a machine with the CLI somewhere
+    /// odd is asked at the same path its sessions run.
+    pub program: String,
 }

 /// Where Claude Code keeps its credentials, as a shell word rather than a path:
@@ -170,7 +194,7 @@ impl ClaudeUsage {

 impl UsageProvider for ClaudeUsage {
    fn name(&self) -> &'static str {
-        "claude"
+        CLAUDE
    }

    fn fetch(&self) -> UsageSnapshot {
@@ -178,36 +202,111 @@ impl UsageProvider for ClaudeUsage {
            Ok(token) => token,
            Err(state) => return self.snapshot(state, Vec::new()),
        };
-        let text = match ureq::get(USAGE_URL)
+        let body = match self.call(&token) {
+            Ok(body) => body,
+            Err(Refused::Other(detail)) => {
+                return self.snapshot(UsageState::Failed { detail }, Vec::new());
+            }
+            Err(Refused::Unauthorized) => match self.after_cli_refresh(&token) {
+                Ok(body) => body,
+                Err(state) => return self.snapshot(state, Vec::new()),
+            },
+        };
+        self.snapshot(UsageState::Ok, parse_windows(&body))
+    }
+}
+
+/// Why one call to the usage endpoint did not produce numbers.
+///
+/// 401 is apart from the rest because it is the only one with a way out: the
+/// endpoint answered, and it means the access token has expired rather than
+/// that anything is broken.
+enum Refused {
+    Unauthorized,
+    Other(String),
+}
+
+impl ClaudeUsage {
+    /// One call to the endpoint with one token.
+    fn call(&self, token: &str) -> Result<Value, Refused> {
+        let text = ureq::get(USAGE_URL)
            .header("Authorization", &format!("Bearer {token}"))
            .header("anthropic-beta", "oauth-2025-04-20")
            .header("User-Agent", USER_AGENT)
            .call()
            .and_then(|mut response| response.body_mut().read_to_string())
-        {
-            Ok(text) => text,
-            Err(err) => {
-                // The error string can embed the URL but never the token.
-                return self.snapshot(
-                    UsageState::Failed {
-                        detail: format!("usage endpoint unreachable: {err}"),
-                    },
-                    Vec::new(),
-                );
-            }
-        };
-        let body: Value = match serde_json::from_str(&text) {
-            Ok(body) => body,
-            Err(err) => {
-                return self.snapshot(
-                    UsageState::Failed {
-                        detail: format!("usage endpoint sent non-JSON: {err}"),
-                    },
-                    Vec::new(),
-                );
-            }
-        };
-        self.snapshot(UsageState::Ok, parse_windows(&body))
+            // The error string can embed the URL but never the token.
+            .map_err(|err| match err {
+                ureq::Error::StatusCode(401) => Refused::Unauthorized,
+                other => Refused::Other(why(&other)),
+            })?;
+        serde_json::from_str(&text)
+            .map_err(|err| Refused::Other(format!("usage endpoint sent non-JSON: {err}")))
+    }
+
+    /// Have the machine's own CLI refresh its token, then ask once more.
+    ///
+    /// **The CLI does the refresh, never this.** Anthropic's OAuth rotates the
+    /// refresh token, so whoever refreshes second presents a dead one and the
+    /// machine is logged out until somebody runs `/login` on it -- and the
+    /// machine we would be refreshing on is usually one with a live session of
+    /// its own. Running the CLI keeps it the only writer of
+    /// `.credentials.json`.
+    ///
+    /// `doctor` rather than the `auth status` it reads like, measured against
+    /// this CLI (2.1.258) on 2026-09-05 with a deliberately invalid token:
+    /// `auth status` reports `loggedIn: true` off the file alone and never
+    /// touches the network, so it would have refreshed nothing while looking
+    /// like it had. `doctor` resolves the account, which is what makes it
+    /// refresh, and it spends no quota. The same probe showed what a *failed*
+    /// refresh does -- the CLI blanks both tokens -- so this must stay on the
+    /// 401 path, where the access token is already dead, and never be used to
+    /// refresh speculatively.
+    ///
+    /// Only a token that actually changed is retried, so a CLI that refreshed
+    /// nothing costs one call rather than two, and this cannot become a loop.
+    fn after_cli_refresh(&self, stale: &str) -> Result<Value, UsageState> {
+        let launch = Launch::new(&self.program, vec!["doctor".to_string()], None);
+        if let Err(err) = self.transport.capture_blocking(&launch) {
+            return Err(UsageState::Failed {
+                detail: format!(
+                    "the Claude login on {} has expired, and `{} doctor` couldn't be run there to refresh it: {err:#}",
+                    self.setup_name, self.program
+                ),
+            });
+        }
+        let fresh = self.access_token()?;
+        if fresh == stale {
+            return Err(self.still_expired());
+        }
+        self.call(&fresh).map_err(|err| match err {
+            Refused::Unauthorized => self.still_expired(),
+            Refused::Other(detail) => UsageState::Failed { detail },
+        })
+    }
+
+    /// A login the CLI could not renew: the one state here somebody has to act
+    /// on, so it says where and what to run.
+    fn still_expired(&self) -> UsageState {
+        UsageState::Failed {
+            detail: format!(
+                "the Claude login on {} has expired and could not be refreshed; run `{} /login` there",
+                self.setup_name, self.program
+            ),
+        }
+    }
+}
+
+/// What a failed call to the usage endpoint should say.
+///
+/// A status is not a network fault and must not be reported as one: the
+/// endpoint answered. 401 never reaches here -- it has its own way out in
+/// [`ClaudeUsage::after_cli_refresh`] -- so what is left is a refusal nobody
+/// on this side can fix.
+fn why(err: &ureq::Error) -> String {
+    match err {
+        ureq::Error::StatusCode(code) => format!("usage endpoint refused the request: HTTP {code}"),
+        other => format!("usage endpoint unreachable: {other}"),
    }
 }

@@ -278,24 +377,260 @@ fn parse_windows(body: &Value) -> Vec<UsageWindow> {
        .collect()
 }

+/// An invented answer, so the screens that draw these can be exercised
+/// without an account.
+///
+/// Every state the usage bar and the usage dialog can be in is otherwise
+/// reachable only by spending somebody's quota or by breaking a machine:
+/// a number near the top, a machine nobody has logged into, one that
+/// cannot be reached, a window between blocks with no reset time. Those
+/// are exactly the states worth looking at, and the ones nobody looks at
+/// because arranging them costs real turns. An echo session sets this
+/// with `/usage` (see `session::echo`), which is the same bargain the
+/// rest of that driver makes: the fixture is invented, what is real is
+/// the path it travels.
+///
+/// Shared by the session layer, which writes it, and [`UsageMonitor`],
+/// which reads it. Empty until something sets it, and an empty fixture
+/// produces no snapshot at all -- an echo session meters nothing, and
+/// nothing is what the phone should draw.
+#[derive(Clone, Default)]
+pub struct Fixture {
+    said: Arc<Mutex<Option<Reported>>>,
+}
+
+/// What a meter answered: which of the four states it is in, and whatever
+/// windows go with it. Empty for every state but [`UsageState::Ok`].
+type Reported = (UsageState, Vec<UsageWindow>);
+
+/// How long the invented five-hour window has left, when nothing says.
+const FIXTURE_MINUTES: i64 = 125;
+
+impl Fixture {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    fn is_set(&self) -> bool {
+        self.said.lock().unwrap().is_some()
+    }
+
+    fn read(&self) -> Option<Reported> {
+        self.said.lock().unwrap().clone()
+    }
+
+    /// Acts on the words typed after `/usage`, and says what it did.
+    ///
+    /// The vocabulary lives here rather than in the echo driver because
+    /// these are this module's states: a driver spelling them out would
+    /// be a second place that has to learn about a fifth one.
+    pub fn command(&self, words: &str) -> String {
+        let mut words = words.split_whitespace();
+        let Some(first) = words.next() else {
+            return match self.read() {
+                Some((state, windows)) => format!("usage fixture: {}", describe(&state, &windows)),
+                None => "usage fixture: unset, so this session meters nothing. \
+                         `/usage 42` puts up a five-hour window at 42%."
+                    .to_string(),
+            };
+        };
+        let rest: Vec<&str> = words.collect();
+        let detail = || {
+            if rest.is_empty() {
+                "set by /usage".to_string()
+            } else {
+                rest.join(" ")
+            }
+        };
+        let (state, windows) = match first {
+            "off" | "none" | "clear" => {
+                *self.said.lock().unwrap() = None;
+                return "usage fixture cleared: this session meters nothing again".to_string();
+            }
+            "notloggedin" | "logged-out" => (UsageState::NotLoggedIn, Vec::new()),
+            "unreachable" => (UsageState::Unreachable { detail: detail() }, Vec::new()),
+            "failed" => (UsageState::Failed { detail: detail() }, Vec::new()),
+            percent => match percent.parse::<f64>() {
+                Ok(percent) => (
+                    UsageState::Ok,
+                    fixture_windows(percent.clamp(0.0, 100.0), rest.first().copied()),
+                ),
+                Err(_) => {
+                    return format!(
+                        "\"{percent}\" is not one of this fixture's answers. Say a percentage \
+                         (`/usage 42`, optionally with `90` minutes left, `never` for a window \
+                         between blocks, or `unreadable` for a reset time that cannot be read), \
+                         or one of `notloggedin`, `unreachable`, `failed`, `off`."
+                    );
+                }
+            },
+        };
+        let said = describe(&state, &windows);
+        *self.said.lock().unwrap() = Some((state, windows));
+        format!("usage fixture set: {said}")
+    }
+}
+
+/// The three windows Claude reports today, invented around one number.
+///
+/// Three rather than one because the bar under a session header reads the
+/// five-hour window and the dialog behind the button draws all of them,
+/// and a fixture with one window leaves half the screen untested. The
+/// weekly ones are derived from the same figure so that the worst of them
+/// -- which is what colours the button -- is still the one asked for.
+fn fixture_windows(percent: f64, reset: Option<&str>) -> Vec<UsageWindow> {
+    let resets_at = match reset {
+        // The state a real response is in between blocks: there is no
+        // window running, so there is nothing to reset. It is not a
+        // missing value, and the phone words it differently.
+        Some("never") | Some("none") => None,
+        // A timestamp that arrives and cannot be read, which is the one
+        // case that really is "we could not find out".
+        Some("unreadable") | Some("bad") => Some("whenever it feels like it".to_string()),
+        other => Some(reset_in(
+            other
+                .and_then(|word| word.parse().ok())
+                .unwrap_or(FIXTURE_MINUTES),
+        )),
+    };
+    vec![
+        UsageWindow {
+            kind: "session".to_string(),
+            label: "5-hour window".to_string(),
+            percent,
+            resets_at: resets_at.clone(),
+            active: true,
+        },
+        UsageWindow {
+            kind: "weekly_all".to_string(),
+            label: "Weekly (all models)".to_string(),
+            percent: percent / 2.0,
+            resets_at: resets_at.as_ref().map(|_| reset_in(FIXTURE_MINUTES * 40)),
+            active: false,
+        },
+        UsageWindow {
+            kind: "weekly_scoped".to_string(),
+            label: "Weekly (Echo)".to_string(),
+            percent: percent / 4.0,
+            resets_at: resets_at.as_ref().map(|_| reset_in(FIXTURE_MINUTES * 40)),
+            active: false,
+        },
+    ]
+}
+
+/// `minutes` from now, in the format the real endpoint sends.
+fn reset_in(minutes: i64) -> String {
+    let at = time::OffsetDateTime::now_utc() + time::Duration::minutes(minutes);
+    at.format(&time::format_description::well_known::Rfc3339)
+        // Formatting a timestamp cannot fail for any input this builds;
+        // saying so beats a fixture that silently has no reset time.
+        .unwrap_or_else(|_| "unformattable".to_string())
+}
+
+/// One line naming what a fixture is currently claiming, for the reply
+/// the echo session writes back.
+fn describe(state: &UsageState, windows: &[UsageWindow]) -> String {
+    match state {
+        UsageState::Ok => match windows.first() {
+            Some(window) => format!(
+                "{}% of the five-hour window, {}",
+                window.percent,
+                match &window.resets_at {
+                    Some(at) => format!("resetting at {at}"),
+                    None => "with no reset time (the between-blocks state)".to_string(),
+                }
+            ),
+            None => "no windows at all".to_string(),
+        },
+        UsageState::NotLoggedIn => "nobody is logged in on this machine".to_string(),
+        UsageState::Unreachable { detail } => format!("machine unreachable ({detail})"),
+        UsageState::Failed { detail } => format!("the meter failed ({detail})"),
+    }
+}
+
+/// The fixture, as a provider, so it travels the same route and the same
+/// cache as a real meter rather than being spliced in at the screen.
+struct EchoUsage {
+    setup: String,
+    setup_name: String,
+    fixture: Fixture,
+}
+
+impl UsageProvider for EchoUsage {
+    fn name(&self) -> &'static str {
+        ECHO
+    }
+
+    fn fetch(&self) -> UsageSnapshot {
+        let (state, windows) = self
+            .fixture
+            .read()
+            // Only ever built for a fixture that is set; a race with
+            // `/usage off` between the two reads lands here, and "the
+            // machine could not be asked" is the honest word for it.
+            .unwrap_or((
+                UsageState::Unreachable {
+                    detail: "the usage fixture was cleared".to_string(),
+                },
+                Vec::new(),
+            ));
+        UsageSnapshot {
+            provider: self.name().to_string(),
+            setup: self.setup.clone(),
+            setup_name: self.setup_name.clone(),
+            state,
+            windows,
+            fetched_at: crate::session::now(),
+        }
+    }
+
+    /// Read from memory, and set by somebody who is about to look at the
+    /// screen it changes.
+    fn poll_interval(&self) -> Duration {
+        Duration::ZERO
+    }
+}
+
 /// Which paid services a machine can be asked about.
 ///
 /// Derived from what the setup says it can run, so a machine with no Claude
 /// provider is not asked about Claude limits -- it has none, and a row saying
-/// so would be a fact about nothing. A second service later adds a branch here
-/// and an impl beside [`ClaudeUsage`], not a screen.
-fn providers_for(setup: &SetupConfig) -> Vec<Box<dyn UsageProvider>> {
+/// so would be a fact about nothing.
+///
+/// Which meter a provider has is [`DriverKind::usage_provider`]'s answer rather
+/// than a second match on kinds here, because the phone pairs a session with
+/// one of these rows by that same name: two lists that disagreed would leave a
+/// session looking for a snapshot nothing produces. A second service later is a
+/// name there and an impl beside [`ClaudeUsage`], not a screen.
+fn providers_for(setup: &SetupConfig, fixture: &Fixture) -> Vec<Box<dyn UsageProvider>> {
    let mut found: Vec<Box<dyn UsageProvider>> = Vec::new();
-    if setup
-        .providers
-        .iter()
-        .any(|provider| provider.kind == DriverKind::ClaudeCli)
-    {
-        found.push(Box::new(ClaudeUsage {
-            setup: setup.id.clone(),
-            setup_name: setup.name.clone(),
-            transport: Transport::for_setup(setup),
-        }));
+    for provider in &setup.providers {
+        let Some(name) = provider.kind.usage_provider() else {
+            continue;
+        };
+        // A machine offering two Claude providers has one account, not
+        // two: the meter belongs to the machine and the service, which is
+        // exactly what the cache is keyed by.
+        if found.iter().any(|already| already.name() == name) {
+            continue;
+        }
+        match name {
+            CLAUDE => found.push(Box::new(ClaudeUsage {
+                setup: setup.id.clone(),
+                setup_name: setup.name.clone(),
+                transport: Transport::for_setup(setup),
+                program: provider.program().to_string(),
+            })),
+            // Nothing at all until a test has asked for something: an
+            // echo session costs nothing, so the honest answer is no row
+            // rather than a row saying zero.
+            ECHO if fixture.is_set() => found.push(Box::new(EchoUsage {
+                setup: setup.id.clone(),
+                setup_name: setup.name.clone(),
+                fixture: fixture.clone(),
+            })),
+            _ => {}
+        }
    }
    found
 }
@@ -313,11 +648,18 @@ type Cached = HashMap<(String, &'static str), (Instant, UsageSnapshot)>;
 /// machine per service per [`MIN_POLL_INTERVAL`], however often the phone asks.
 pub struct UsageMonitor {
    cache: Mutex<Cached>,
+    /// The invented meter an echo session can put up; empty unless one
+    /// has. Shared with the session layer, which is where the command
+    /// that sets it is typed -- see [`Fixture`].
+    fixture: Fixture,
 }

 impl UsageMonitor {
-    pub fn new() -> Self {
-        Self::default()
+    pub fn new(fixture: Fixture) -> Self {
+        Self {
+            cache: Mutex::new(Cached::new()),
+            fixture,
+        }
    }

    /// One snapshot per machine that offers a paid service, in the order the
@@ -329,10 +671,10 @@ impl UsageMonitor {
    pub fn snapshots(&self, setups: &[SetupConfig]) -> Vec<UsageSnapshot> {
        let mut fresh = Vec::new();
        for setup in setups {
-            for provider in providers_for(setup) {
+            for provider in providers_for(setup, &self.fixture) {
                let key = (setup.id.clone(), provider.name());
                if let Some((fetched, snapshot)) = self.cache.lock().unwrap().get(&key)
-                    && fetched.elapsed() < MIN_POLL_INTERVAL
+                    && fetched.elapsed() < provider.poll_interval()
                {
                    // Cached numbers, but the machine's *name* is read fresh: a
                    // rename should show immediately rather than waiting out a
@@ -367,6 +709,44 @@ impl UsageMonitor {
 #[cfg(test)]
 mod tests {
    use super::*;
+    use crate::config::DriverKind;
+
+    #[test]
+    fn a_refusal_the_endpoint_answered_is_not_reported_as_an_unreachable_one() {
+        assert!(why(&ureq::Error::StatusCode(500)).contains("HTTP 500"));
+        assert!(why(&ureq::Error::HostNotFound).contains("unreachable"));
+    }
+
+    #[test]
+    fn an_expired_login_says_where_to_log_in_rather_than_naming_the_network() {
+        let provider = ClaudeUsage {
+            setup: "far".to_string(),
+            setup_name: "somewhere else".to_string(),
+            transport: Transport::for_setup(&unreachable_setup()),
+            program: "/opt/claude".to_string(),
+        };
+        // The machine cannot be reached, so the refresh attempt fails there
+        // rather than at the endpoint -- and the message still has to name the
+        // machine and the command, since that is all anybody gets to act on.
+        let UsageState::Failed { detail } = provider
+            .after_cli_refresh("stale")
+            .expect_err("an unreachable machine cannot refresh anything")
+        else {
+            panic!("an expired login is a fault to report, not a logged-out machine");
+        };
+        assert!(detail.contains("somewhere else"), "{detail}");
+        assert!(detail.contains("/opt/claude doctor"), "{detail}");
+        assert!(
+            !detail.contains("stale"),
+            "the token must never be quoted back"
+        );
+
+        let UsageState::Failed { detail } = provider.still_expired() else {
+            panic!("still expired is a fault");
+        };
+        assert!(detail.contains("/opt/claude /login"), "{detail}");
+        assert!(!detail.contains("unreachable"), "{detail}");
+    }

    #[test]
    fn parses_the_limits_array_defensively() {
@@ -404,6 +784,7 @@ mod tests {
                port: None,
                identity_file: None,
                options: vec!["ConnectTimeout=1".to_string()],
+                models_dir: None,
                attachments_dir: None,
            }),
            providers: vec![crate::config::ProviderConfig {
@@ -421,6 +802,7 @@ mod tests {
            setup: "far".to_string(),
            setup_name: "somewhere else".to_string(),
            transport: Transport::for_setup(&unreachable_setup()),
+            program: "claude".to_string(),
        };
        let snapshot = provider.fetch();
        // The distinction the old single `error` string could not make: this
@@ -475,9 +857,62 @@ mod tests {
            models: vec![],
        }];
        // A machine with no Claude on it has no Claude limits, and a row
-        // reporting on it would be a fact about nothing.
-        assert!(providers_for(&echo_only).is_empty());
-        assert_eq!(providers_for(&unreachable_setup()).len(), 1);
+        // reporting on it would be a fact about nothing. Echo included:
+        // an echo session spends nothing, so until a fixture says
+        // otherwise there is no meter to report.
+        let unset = Fixture::new();
+        assert!(providers_for(&echo_only, &unset).is_empty());
+        assert_eq!(providers_for(&unreachable_setup(), &unset).len(), 1);
+
+        // And with one set, that machine has exactly the invented meter
+        // -- under the name the session's `usageProvider` will name.
+        let fixture = Fixture::new();
+        fixture.command("42");
+        let found = providers_for(&echo_only, &fixture);
+        assert_eq!(found.len(), 1);
+        assert_eq!(found[0].name(), ECHO);
+        assert_eq!(DriverKind::Echo.usage_provider(), Some(ECHO));
+        assert_eq!(DriverKind::ClaudeCli.usage_provider(), Some(CLAUDE));
+        // A local model costs nothing to run, so it meters nothing.
+        assert_eq!(DriverKind::LlamaCpp.usage_provider(), None);
+    }
+
+    /// The states the fixture exists to make reachable, and the one thing
+    /// it must not do: invent a reset time for a window that has none.
+    #[test]
+    fn the_fixture_says_each_state_the_screens_have_to_draw() {
+        let fixture = Fixture::new();
+        assert!(fixture.read().is_none(), "unset until somebody sets it");
+
+        fixture.command("42 90");
+        let (state, windows) = fixture.read().expect("set");
+        assert_eq!(state, UsageState::Ok);
+        assert_eq!(windows[0].kind, "session");
+        assert_eq!(windows[0].percent, 42.0);
+        assert!(windows[0].resets_at.is_some());
+
+        // Between blocks: no reset time, which the phone words as the
+        // window not running rather than as a time it could not read.
+        fixture.command("42 never");
+        assert_eq!(fixture.read().expect("set").1[0].resets_at, None);
+
+        fixture.command("unreachable no route to host");
+        assert!(matches!(
+            fixture.read().expect("set").0,
+            UsageState::Unreachable { detail } if detail == "no route to host",
+        ));
+
+        fixture.command("off");
+        assert!(fixture.read().is_none());
+
+        // A word it does not know changes nothing and says what it takes.
+        fixture.command("42");
+        let refused = fixture.command("sideways");
+        assert!(
+            refused.contains("not one of this fixture's answers"),
+            "{refused}"
+        );
+        assert_eq!(fixture.read().expect("still set").1[0].percent, 42.0);
    }

    #[test]