Five things asked for on the phone, and one trap behind the first of them.
A llama prompt carries `<__media__>` where a picture was, and llama.cpp
pairs each marker with a decoded image when it tokenizes -- so a marker in
words nobody attached a picture to fails the turn, and then fails every
later one, since the conversation is folded out of a transcript that holds
it for good. A model saying the marker back is enough to do it. Every
message with words in it now goes through `without_marker`.
A model's `general.name` is filled in by whatever converted the file, and
`convert_hf_to_gguf.py` fills it from the directory it converted: Prism ML's
Bonsai publishes `general.name = "Hf"`, which is unique and so passed the
label cascade and told a reader nothing. A name is now used only where it
shares a word with the repo or the file it came from; one that does not
drops to the file name.
The model settings dialog and the server card said what a save would cost in
a paragraph under the control, read after the decision if at all. Both ask
instead, in the shape the rest of that screen already uses for Stop and
Delete -- and the dialog's question is asked over the edits, so Cancel comes
back to them.
A field's label was `labelMedium` in the variant colour while every setting
beside it was body text, which on one form read as two ranks of setting.
`GET /sessions/{id}`'s `transcriptFile` now carries the file's size, and the
settings screen draws it beside what this phone has cached.
Checked on the emulator against the sandbox: the labels line up, the row
says "14 kB · 14 kB cached", and both confirmations appear over a loaded
Qwen3-0.6B. `cargo test` 275 passed, clippy and fmt clean, lint clean.
The Models tab was about this backend's own disk, which is the wrong disk
for every session that runs anywhere else: llama.cpp reads the file where
it runs. So the models of a machine live under that machine's llama.cpp
provider now, beside the settings deciding how each is loaded, and the
download that produces one happens there.
A download is a detached `curl` on that machine, started by a script this
server writes and never spoken to again. Its state is a file beside the
partial, so nothing about it is held here: it survives the app closing,
this backend restarting and a second device watching, and the progress is
`wc -c` of the partial against the size HuggingFace published rather than
anything remembered. A run whose process is gone is reported failed, since
`kill -0` is asked at each listing, and there is no "finished" state -- a
download that finished is a model, in the list beside the ones still
going. Resuming is guarded by the published sha256, which is also checked
before the file takes its real name.
Two other things the same screens wanted:
A provider is drawn as a card rather than as a line of text, bordered
against the machine card it sits in -- the tint it had was one step along
the surface ladder and rendered as one flat block -- with room to tap and
no chevron.
Nothing in a raw block wraps any more; the block scrolls sideways
instead, one offset for all its lines, so a diff or a column-aligned test
run still reads as one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A llama.cpp session had its own `llama-server`: two sessions on one model
held two copies of it in memory, a model change bought a load only that
session benefited from, and the process was a session's to end. A machine's
models are now served by one `llama-server` in **router mode** -- no `-m`,
a preset file naming models and their flags, a child server per model asked
for, and each request routed by its `model` field. So one server per model
with that model's own settings is what a machine runs, while this backend
has one process, one port and one record per machine to keep track of.
The record is the mechanism every other driver already uses, so a restart
adopts it; a session records the same pid in its own directory as
`Detail::Shared`, and `process::signal` refuses to signal one of those --
which is what keeps stopping, deleting or cleaning up after one session
from unloading a model every other session is using. Nothing stops a router
on its own. That is deliberate (a loaded model is minutes of disk) and it is
why the machines tab now has a card per provider that opens its own screen:
how each model is loaded, how many stay in memory, Unload, and Stop.
How a model is *loaded* therefore belongs to the model on its machine rather
than to a session -- context size, GPU layers, threads, slots, speculative
decoding -- written into the preset as llama-server's own argument names.
Saving them re-reads that file, which unloads the model; that is the change
taking effect, and the dialog says so before you save. What stays a
session's is everything that rides on a request, including which tools it
offers: the router hosts one set for the machine and the choice is a filter
applied here, so it costs no reload (2,181 tokens of prompt with all seven,
698 with none).
Verified end to end against the scratch backend and the emulator: two
sessions sharing one loaded model with one child process, a second session
joining it with a 26ms prefill, a backend restart adopting the router and
answering with the prompt cache intact, the same over ssh to this VM, a
model's settings reaching the running server, Unload, and Stop leaving every
session `exited` with no error line.
A llama session was a chat box: no tools, a fixed model, no permission
mode, and a model name drawn as the path the file sits at. It now runs the
agent loop itself, which is what the pieces below all hang off.
Tools are `llama-server`'s own (`--tools all`), which that server both
publishes and runs -- `GET /tools` for the definitions, `POST /tools` to
call one. Web search is Exa's MCP server, reached from this backend rather
than from the machine serving the model: that is what llama.cpp's own web
UI does, and it puts the search on the machine with a route out instead of
the one with the GPU. `llama-server`'s `--mcp-servers-json` can only spawn
local commands, so using it would have meant a Node bridge on every
machine that serves a model.
Driving the loop is what makes the permission gate ours. Two modes,
`manual` and `bypassPermissions`, which is what the mechanism has: the web
UI asks before every call and remembers the tools you say "always" to. The
allowances fold back out of the transcript's own answers, so they survive
a restart and a model change without being stored anywhere else.
Also here, because tools made each of them matter:
- **Loading is a state.** A 12 GB model takes twenty seconds to reach
memory and refuses everything until it has; the session used to report
`running` for that whole time, and a message sent meanwhile came back as
an error. It is `loading` now, and the message waits.
- **The model can be changed.** A `llama-server` holds one model, so this
stops it and starts another. The conversation survives because it was
never in the server.
- **Models are named, not pathed.** `general.name` read out of the file
itself -- over ssh too, in the round trip the spawn was already making.
Where two models share a name the file name breaks the tie.
- **`-np 1`, and the MTP draft head where the file has one.** Measured on
the 27B here: 41.5 tok/s plain, 61.4 with `--spec-type draft-mtp` at one
slot, and 28 with it at four -- speculating against a split KV cache is
worse than not speculating. The flag is conditional because asking for a
head that is not there makes `llama-server` exit.
- **A refusal says what to do.** Tool results are thousands of tokens, so
an overrun context is now ordinary; it was "http status: 400" and is now
the server's own "exceeds the available context size, try increasing it".
`GET /machines/{id}/models` is gone: the provider models route answers the
same question, and two answers to one question is how a picker comes to
offer a model the spawn screen does not.
Verified end to end against real models: a tool call asked and allowed, an
Exa search, a shell command, a 27B loaded while a message waited on it, a
model switch mid-session, a second message queued behind a running turn,
and the whole of it again on a session running over ssh.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The rate-limit bar answered a question about an account, and picked the
answer by machine. One machine runs echo, the Claude CLI and a local
model side by side, so every echo session on it drew the CLI's five-hour
window: a quota that session cannot spend and could never run down. A
session now names its meter (`usageProvider`, from
`DriverKind::usage_provider`, which `usage::providers_for` reads too so
the two lists cannot disagree), and the phone matches on machine *and*
provider. Nothing meters echo or llama, and nothing at all is drawn --
including while the first fetch is out, since "checking" under a session
that turns out to meter nothing is a row the screen then withdraws.
Echo gets a meter it can be *told* about instead: `/usage 42`,
`/usage 95 20`, `/usage 42 never`, `/usage notloggedin`,
`/usage unreachable`, `/usage failed`, `/usage off`. Those states cost
real quota to arrange, which is why none of them had been looked at.
And llama.cpp runs wherever a setup says, which was the last of phase 5.
`Transport::reserve_port` is the second half of what a transport is --
"run this" plus "reach this port" -- returning the port the server binds
there and the port that reaches it here, and `Launch::reaching` puts the
`-L` tunnel on the connection that already carries the command. Three
things that came out of building it:
- A forwarded launch gets a pty and every other one keeps `-T`. Killing
the ssh client ends a CLI by closing the stdin it reads; 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.
- The model is looked for on the machine that will serve it, at that
machine's own models directory, so `GET /setups/{id}/models` is what
the spawn screen offers rather than the backend's own downloads.
- The readiness poll watches the process, not only the port: a model
that will not load exits in a second and would otherwise have been
reported as "gave up after 300s". The failure carries the log's tail.
Exercised end to end against this VM over ssh to itself: spawn, load,
answer, outlive a backend restart, be adopted, answer again, and stop --
with both the ssh client and the far llama-server gone afterwards. The
local path, the Claude bar and the spawn screen checked on the emulator.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The markdown had accumulated a lot that was stale rather than wrong.
PLAN.md still described pi as the llama.cpp harness, a refcounted
LlamaServerManager, and a providers-by-hosts cross-product, all of which
were superseded or never built; it also carried a second copy of the HTTP
table that routes.rs owns. EXPLORER.md and TRANSCRIPT_CACHE.md held
implementation checklists for work that has since landed. AGENTS.md
restated most of PLAN.md's design instead of being the working-notes
layer it says it is. 3225 lines of markdown to 2180, with the stale
sections gone rather than reworded.
On the server, comments explaining what the code already says are out and
the ones recording a constraint, a measurement or an incident are kept but
cut to a few lines each: 5504 comment lines to 4586.
Four doc comments in session/mod.rs, and one each in process.rs and
usage.rs, had drifted onto the item above the one they describe --
functions were reordered without them, so `stop_session`'s doc sat on
`set_session_cwd`, `stat_of`'s on `struct Stat`, and `UsageMonitor`'s on
`type Cached`. Each is back on its own item.
routes.rs's module table also claimed later phases would add `/hosts`,
which setups replaced.
cargo test (127 passed), clippy --all-targets and fmt are clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A Rust backend that owns the sessions and an Android app that reads them.
The server spawns and adopts CLI processes, normalises everything they emit
into one event model, keeps the transcript, and serves it over pinned TLS on
a WireGuard interface; the phone streams that, replies, sends images, and
imports conversations the machine already has.
`AGENTS.md` is the working guide -- what runs where, what has been measured,
and the faults that were expensive to find. `PLAN.md` is the design record.
History before this point was squashed away. It was a personal project's
running commentary and carried a name and a couple of machine paths that
have no business in a public repository; the tree is what mattered and the
tree is here.
iris-ai
iris