Give llama.cpp sessions tools, web search and a model picker

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>

.claude/skills/ai-app-rigs/SKILL.md

@@ -143,18 +143,42 @@ 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 `$ORIGIN`, so no `LD_LIBRARY_PATH` is
-needed. One model is downloaded — `unsloth/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.gguf`,
-639 MB under `~/.local/share/ai-app/models` — and answers at usable speed on
-this VM's 8 cores. **Do not test with a 2-bit quant**: the
-IQ2_XXS of that model produces fluent nonsense, which reads exactly like a
-broken driver — `llama-cli` produces the same from the file directly, which
-is how to tell the two apart in a hurry.
+**Both are set up here** and need nothing typed. The prebuilt llama.cpp lives
+outside the repo at `~/.local/opt/llama.cpp-vk` — a **Vulkan** build as of
+2026-09-19, replacing the CPU one that was there before — and is symlinked as
+both `~/.local/bin/llama-server` and `/usr/local/bin/llama-server`. The second
+is what makes **discovery find it over ssh**: `~/.local/bin` is not on the
+PATH a non-interactive ssh session gets. It resolves its own libraries through
+`$ORIGIN`, so no `LD_LIBRARY_PATH` is needed.
+
+Two models are downloaded under `~/.local/share/ai-app/models`:
+
+- `unsloth/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.gguf`, 639 MB, loads in ~4s. It
+  calls tools correctly and is the right rig for the driver's shape. Do not
+  judge *answers* by it — asked for the second line of a file it read from
+  line 2 and then named the third.
+- `ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF/Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp.gguf`,
+  12 GB, ~20s to load, and the only one here with a multi-token-prediction
+  head. It is the rig for anything about `loading` being a state of its own,
+  since 20s is long enough to send into.
+
+**Do not test with a 2-bit quant**: the IQ2_XXS of the 0.6B 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.
+
+**The GPU is shared and llama-server dies loudly when it runs out.** A second
+server loading a model while the 27B holds VRAM fails with `radv/amdgpu:
+Failed to allocate a buffer` / `MESA: error: buffer allocation failed` and
+exits mid-request. `-ngl 0` runs it on the 8 cores instead, which is the way
+to test the driver while something else holds the card.
+
+**Testing tools and MCP without the app**: `llama-server --tools all` publishes
+its built-in tools at `GET /tools` and runs one at `POST /tools` with
+`{"tool": …, "params": …}` and an `x-tool-cwd` header — so a whole agent loop
+is drivable with `curl` and no model at all. The Exa MCP server at
+`https://mcp.exa.ai/mcp` answers **without an API key** and needs a
+`User-Agent` header (Cloudflare answers 403 without one, which reads as a
+refusal rather than a missing header).
 
 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
@@ -212,6 +236,30 @@ where it was instead of half-deleted.
 
 ## Measurements worth not re-taking
 
+- **`-np 1` is what makes the MTP draft head pay.** Taken 2026-09-19 on the
+  27B above, decode speed for a 300-token reply, from `llama-server`'s own
+  timings rather than the clock:
+
+  | flags | tok/s |
+  | --- | --- |
+  | plain, any `-np` | 41.5 |
+  | `--spec-type draft-mtp -np 1` | 61.4 |
+  | `--spec-type draft-mtp -np 2` (n-max 2) | 65.9 |
+  | `--spec-type draft-mtp`, default `-np` (4 slots) | 28 |
+
+  Draft acceptance is 0.53–0.73 in every case, so the head is working in all
+  of them: what changes is that speculating against a KV cache split four ways
+  is slower than not speculating. The driver passes `-np 1` always, so this is
+  recorded for whoever next sees MTP look broken. `--spec-draft-n-max 2` was
+  worth another 7% in a single sample and is deliberately *not* passed — one
+  sample on a virtualised GPU is not a number to hardcode.
+
+- **Asking for the head when the file has none is fatal**, not ignored:
+  `context type MTP requested but model doesn't contain MTP layers` and the
+  server exits. Without the flag the same file logs `unused tensor
+  blk.N.nextn.* — ignoring` and runs normally, which is the state to look for
+  when MTP is silently not happening.
+
 - **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

AGENTS.md

@@ -47,6 +47,21 @@ Module-by-module intent is in PLAN.md's "Backend layout".
   readiness poll watches the process as well as the port, since a model that
   will not load exits in a second and was being reported as "gave up after
   300s". See PLAN.md's "Transport" and "llama-server management".
+  **A llama session has tools and runs the loop itself** (2026-09-19):
+  `--tools all` gives it `llama-server`'s built-in set, which that server also
+  *runs* (`GET /tools` for the definitions, `POST /tools` to call one), while
+  web search comes from an MCP server this backend connects to directly
+  (`session/llama/mcp.rs`, Exa preset in a discovered provider's
+  `mcpServers`). Driving the loop is what makes the permission gate ours:
+  `manual` asks before every call and remembers a tool you answer
+  "Always allow …" to, `bypassPermissions` never asks, and the allowances are
+  folded back out of the transcript. Three more things fall out of it and are
+  easy to get wrong again — a model change **reloads the server** rather than
+  being refused, since the conversation lives in the transcript rather than in
+  `llama-server`; `-np 1` is always passed, and it is what decides whether the
+  MTP draft head is a 50% speed-up or a 33% loss; and `--spec-type draft-mtp`
+  is conditional on the file actually having a head, because asking for one
+  that is not there makes `llama-server` **exit**.
   Codex is one persistent `codex app-server --stdio` process per session; its
   driver uses native turn steering and interruption, persists the protocol
   state and thread id, and reads subscription limits through the same CLI
@@ -310,6 +325,15 @@ written, and the fold uses that same predicate to decide a reply is settled.
 
 ## Things that have bitten
 
+- **A llama session reports `loading`, and a message sent into it waits.**
+  Before 2026-09-19 the session showed `running` from the moment the process
+  started, so a minute of reading a model off disk was indistinguishable from
+  a minute of thinking -- and anything sent in that window came back as an
+  error, because `llama-server` refuses everything until the model is in
+  memory. `SessionStatus::Loading` is the state and `Shared::await_ready` is
+  the waiting. A driver that reports `Loading` owes the holding as well as the
+  word.
+
 - **A transcript outlives the enum.** Removing `Event::TaskNote` hours after
   adding it made every transcript that had recorded one unreadable, so
   `launch` failed for those sessions and `SessionManager::new` skipped them —

PLAN.md

@@ -345,6 +345,55 @@ deliberate and easy to undo by accident:
   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.
+- **Loading is a state of its own** (2026-09-19, `SessionStatus::Loading`).
+  A multi-gigabyte model takes tens of seconds to reach memory and refuses
+  everything until it has, and the session used to report `running` for that
+  whole time — indistinguishable from a model thinking, with the added
+  detail that any message sent meanwhile came back as an error. It is now
+  `loading` on both screens, and a message sent into a load **waits** for it
+  rather than failing. The waiting is the driver's (`Shared::await_ready`, a
+  condvar on a three-state `Serving`), because "there is a process and it is
+  not ready" is a fact only a driver can have. The third state matters as
+  much as the first two: a model that will never load has to answer a waiting
+  message with what went wrong rather than holding it for ever.
+- **The driver runs the agent loop, and therefore owns the permission gate**
+  (2026-09-19). `llama-server --tools all` *hosts* the built-in tools —
+  `GET /tools` is their definitions, `POST /tools` runs one — but it does not
+  drive a conversation: a completion comes back with tool calls in it and
+  stops. So the loop is here, which is what puts "may I run this?" somewhere
+  a phone can answer it. Two modes, `manual` and `bypassPermissions`, which
+  is what the mechanism actually has: llama.cpp's own web UI asks before
+  every call and remembers the tools you said "always" to, and a third mode
+  between them would have to invent a rule about which tools count as edits.
+  The allowances are folded out of the transcript's `Answered` events, like
+  everything else this driver remembers, which is why the answer carries the
+  tool's name in it.
+- **Tools run where the model does; MCP runs here** (2026-09-19). The
+  built-in tools are the far machine's, for the same reason the model file
+  is — they act on that machine's disk. An MCP server is reached from *this*
+  backend instead (`session/llama/mcp.rs`), which is both what llama.cpp's
+  own web UI does (it connects to `https://mcp.exa.ai/mcp` from the browser)
+  and the right side to be on: a web search wants the machine with a route
+  out, not the machine with the GPU. `llama-server`'s own `--mcp-servers-json`
+  is deliberately not used — it can only spawn local commands, so a remote
+  server would mean a Node bridge on whichever machine serves the model.
+- **A model change reloads the server rather than being refused** (2026-09-19).
+  A `llama-server` holds one model, so switching stops it and starts another;
+  the conversation survives because the conversation was never in the server.
+  What is lost is the prompt cache, which is exactly what the phone already
+  warns about before a switch.
+- **One slot, and the draft head where the file has one** (measured
+  2026-09-19). `-np 1` always: a session is one conversation making one
+  request at a time, so the other three slots `llama-server` picks on its own
+  are context this session could have had. It is also what decides whether
+  multi-token prediction pays — on the 27B here, **41.5 tok/s** plain at any
+  slot count, **61.4** with `--spec-type draft-mtp` at one slot, and **28**
+  with the head at four. Speculating against a split KV cache is worse than
+  not speculating, and it reads exactly like the head being broken.
+  The flag is conditional because it must be: asked for on a model without a
+  head, `llama-server` exits. `crate::gguf::has_mtp_head` reads the answer out
+  of the file — on the machine that will serve it, in the round trip the spawn
+  was already making — and `params["speculative"] = "off"` is the way out.
 
 ### Models (2026-08-28)
 
@@ -1386,8 +1435,10 @@ verified by running it, matching dev-updater's posture.
   it truncates old KV cache entries, which is silent forgetting with no
   summary, and it corrupts the harness's view of what the model knows. Fine
   as a server-side safety net; not memory management.
-- **Remote llama-server** needs its port forwarded (`ssh -L`) and is not
-  built; such a session is refused rather than misdirected.
+- **MCP servers are configured in `config.ron`, not from the phone**
+  (2026-09-19). `mcpServers` on a llama provider, with Exa preset on a newly
+  discovered one. A phone screen for them is the obvious next step and was
+  deliberately left out of the change that added them.
 - **Claude sessions over ssh need the remote machine logged in to Claude.**
   Usage reporting reads each machine's own credentials, and the Machines tab
   can run that machine's CLI login without requiring an interactive SSH shell.

app/androidApp/src/main/kotlin/com/example/aiapp/Api.kt

@@ -1334,7 +1334,18 @@ fun deleteSession(settings: ServerSettings, sessionId: String, deleteForeign: Bo
 // Browsing is proxied by the server rather than done here, because this app trusts exactly one
 // certificate and has no general internet trust to spend on huggingface.co.
 
-data class LocalModel(val key: String, val repo: String, val file: String, val bytes: Long)
+data class LocalModel(
+    val key: String,
+    val repo: String,
+    val file: String,
+    val bytes: Long,
+    /**
+     * What the file itself says it is called, or null when it does not say. Not what to draw: see
+     * the server's `models::labels`, which needs the whole list to decide -- two quantisations of
+     * one model share a name.
+     */
+    val name: String?,
+)
 
 /**
  * A download in flight or finished. [total] is null when the server never said how big the file is
@@ -1371,36 +1382,28 @@ private fun parseDownload(o: JSONObject) =
     )
 
 /**
- * The models on one machine, which is the list a llama.cpp session there can choose from.
+ * One model a picker can offer.
  *
- * 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.
+ * Two fields because for one provider they differ: a llama.cpp session names its model by the path
+ * it lives at and reads it as the name its own metadata gives it. Every other provider's [id] is
+ * already what a person calls it, and the server says so by repeating it -- which is what keeps
+ * every picker here free of a branch on the session kind.
  */
-fun fetchMachineModels(settings: ServerSettings, machineId: String): List<LocalModel> =
-    requestFromServer(settings, "/machines/${machineId.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"),
-            )
-        }
-    }
+data class OfferedModel(val id: String, val label: String)
 
-/** The current model catalog for one CLI provider on the machine where it runs. */
+/** The current model catalog for one provider on the machine where it runs. */
 fun fetchProviderModels(
     settings: ServerSettings,
     machineId: String,
     provider: String,
-): List<String> =
+): List<OfferedModel> =
     requestFromServer(
         settings,
         "/machines/${machineId.urlEncoded()}/providers/${provider.urlEncoded()}/models",
     ) { connection ->
-        JSONArray(connection.inputStream.bufferedReader().readText()).strings()
+        JSONArray(connection.inputStream.bufferedReader().readText()).mapObjects { m ->
+            OfferedModel(id = m.getString("id"), label = m.getString("label"))
+        }
     }
 
 fun fetchModels(settings: ServerSettings): Models =
@@ -1414,6 +1417,7 @@ fun fetchModels(settings: ServerSettings): Models =
                         repo = m.getString("repo"),
                         file = m.getString("file"),
                         bytes = m.getLong("bytes"),
+                        name = m.optString("name").ifEmpty { null },
                     )
                 },
             downloads = body.getJSONArray("downloads").mapObjects(::parseDownload),

app/androidApp/src/main/kotlin/com/example/aiapp/Events.kt

@@ -325,7 +325,8 @@ fun parseSeqEvent(json: String): SeqEvent {
  * first time the server grows a state, and the drift would be a reply that never splits or one
  * split mid-stream.
  */
-fun sessionWorking(state: String): Boolean = state == "running" || state == "compacting"
+fun sessionWorking(state: String): Boolean =
+    state == "running" || state == "compacting" || state == "loading"
 
 /** Whether the latest events still say this session needs an explicit provider login. */
 internal fun authenticationPromptAfter(open: Boolean, event: SessionEvent): Boolean =

app/androidApp/src/main/kotlin/com/example/aiapp/ModelName.kt

@@ -21,12 +21,25 @@ const val DEFAULT_MODEL = "default"
  * one model rather than one model from another. Anything that does not look like that is returned
  * untouched.
  *
+ * A llama.cpp session's model is not an identifier at all -- it is `owner/repo/file.gguf`, where
+ * the file was downloaded from -- so what is kept is the file, which is the part that tells two
+ * models apart, and the extension goes with the directories. The model's *own* name is better still
+ * and is not derivable here: it is inside the file, and only the server has ever opened it. Where a
+ * screen has the server's answer it should prefer it; this is the floor under every screen that
+ * does not.
+ *
  * A display decision, not a correction: the full name is what the session reports.
  */
 fun modelLabel(model: String?): String {
     val name = model?.takeIf { it.isNotBlank() } ?: return DEFAULT_MODEL
+    if (name.endsWith(GGUF)) {
+        return name.substringAfterLast('/').removeSuffix(GGUF)
+    }
     return name.removePrefix("claude-").replace(DATED_SUFFIX, "")
 }
 
 /** A trailing `-YYYYMMDD`, which is how these identifiers carry their release date. */
 private val DATED_SUFFIX = Regex("""-\d{8}$""")
+
+/** What every model a llama.cpp session can run is stored as. */
+private const val GGUF = ".gguf"

app/androidApp/src/main/kotlin/com/example/aiapp/SessionScreen.kt

@@ -302,7 +302,7 @@ fun SessionScreen(
     var permissionMode by remember { mutableStateOf(summary.permissionMode ?: "auto") }
     // The models this provider actually offers, asked of the server rather than listed here: a
     // hardcoded list is a claim about a machine.
-    var offeredModels by remember { mutableStateOf<List<String>>(emptyList()) }
+    var offeredModels by remember { mutableStateOf<List<OfferedModel>>(emptyList()) }
     var offeredPermissionModes by remember { mutableStateOf<List<String>>(emptyList()) }
     val lifecycleOwner = LocalLifecycleOwner.current
     // The resume cursor, written from the stream's IO thread.
@@ -1194,6 +1194,17 @@ fun SessionScreen(
             }
     }
 
+    /**
+     * What to call a model on screen.
+     *
+     * The provider's own answer where it has one, because only the server can have it: a llama
+     * model is identified by the path it lives at and named by what is written inside the file, and
+     * the phone has never opened that file. [modelLabel] is the fallback and the right one for the
+     * rest -- a coding CLI's identifier already is its name.
+     */
+    fun label(id: String?): String =
+        offeredModels.firstOrNull { it.id == id }?.label ?: modelLabel(id)
+
     // Only for the model picker, which a subagent does not have.
     if (!isSubagent) {
         LaunchedEffect(summary.machine, summary.provider) {
@@ -1889,8 +1900,8 @@ fun SessionScreen(
         ) {
             pendingModel?.let { chosen ->
                 ModelSwitchWarning(
-                    from = modelLabel(model),
-                    to = modelLabel(chosen),
+                    from = label(model),
+                    to = label(chosen),
                     onDismiss = { pendingModel = null },
                     onConfirm = {
                         pendingModel = null
@@ -2013,22 +2024,28 @@ fun SessionScreen(
                         ) {
                             if (offeredModels.isNotEmpty()) {
                                 PickerButton(
-                                    current = modelLabel(model),
+                                    current = label(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,
+                                    options =
+                                        listOf(DEFAULT_MODEL) + offeredModels.map { it.label },
                                     // 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 ->
+                                    onPick = { picked ->
+                                        // Back to the id, because that is what the server resolves
+                                        // and it is not always the word on the chip.
+                                        val chosen =
+                                            offeredModels.firstOrNull { it.label == picked }?.id
+                                                ?: picked
                                         if (
-                                            modelLabel(chosen) == modelLabel(model) ||
+                                            label(chosen) == label(model) ||
                                                 !worthWarningAbout(status, contextTokens, items)
                                         ) {
                                             act { setSessionModel(settings, summary.id, chosen) }

app/androidApp/src/main/kotlin/com/example/aiapp/SessionStatusWords.kt

@@ -22,6 +22,10 @@ fun sessionStatusWord(status: String, subagent: Boolean = false): String =
         "idle" -> "idle"
         "running" -> "running"
         "compacting" -> "compacting"
+        // Not "running": a model coming off disk is not a model answering, and the difference is
+        // minutes. Said in its own word so a first message that waits is explained rather than
+        // looking like a session that has stopped responding. See `SessionStatus::Loading`.
+        "loading" -> "loading"
         // Its own word, because the state it is easily mistaken for means the opposite: "idle"
         // invites the reader to type something, and a waiting session is going to carry on without
         // them. See `SessionStatus::Waiting`.
@@ -53,6 +57,9 @@ fun sessionStatusColour(status: String): Color =
         "awaitingInput" -> awaitingColor
         "running" -> runningColor
         "compacting" -> commandColor
+        // The same accent as the other states that are busy on their own account, because that is
+        // what this is: something is happening and nothing is wanted from the reader.
+        "loading" -> commandColor
         "waiting" -> waitingColor
         else -> MaterialTheme.colorScheme.onSurfaceVariant
     }

app/androidApp/src/main/kotlin/com/example/aiapp/SpawnScreen.kt

@@ -58,7 +58,7 @@ fun SpawnScreen(
     var providerName by remember { mutableStateOf<String?>(null) }
     var title by remember { mutableStateOf("") }
     var model by remember { mutableStateOf("") }
-    var providerModels by remember { mutableStateOf<List<String>>(emptyList()) }
+    var providerModels by remember { mutableStateOf<List<OfferedModel>>(emptyList()) }
     var providerModelsLoading by remember { mutableStateOf(false) }
     var providerModelsError by remember { mutableStateOf<String?>(null) }
     var cwd by remember { mutableStateOf("") }
@@ -73,12 +73,13 @@ fun SpawnScreen(
     // 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) }
-    // GGUFs on the chosen machine, for a llama provider to choose between. Kept separate from a
-    // coding CLI's provider catalog and refetched when the machine changes.
-    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("") }
+    // Whether a model that carries a multi-token-prediction head drafts with it. Left to the
+    // server by default, which turns it on exactly where the file has one -- see `SPECULATIVE` in
+    // the llama driver. Here so a machine where drafting does not pay has a way out that is not
+    // an edit to config.ron.
+    var speculative by remember { mutableStateOf(SPECULATIVE_AUTO) }
 
     LaunchedEffect(Unit) {
         // Separate from the machines fetch below and deliberately not fatal: failing to learn the
@@ -126,17 +127,6 @@ fun SpawnScreen(
                 is LoadState.Loaded -> state.value
             }
         val machine = machines.firstOrNull { it.name == machineName }
-        // 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(machine?.id) {
-            models = emptyList()
-            modelKey = null
-            val id = machine?.id ?: return@LaunchedEffect
-            models =
-                runCatching { withContext(Dispatchers.IO) { fetchMachineModels(settings, id) } }
-                    .getOrDefault(emptyList())
-        }
         val current = machine?.providers?.firstOrNull { it.name == providerName }
         // Coding CLIs take a working directory, model, permission mode and thinking level. Keying
         // the extra fields on the kind rather than the provider name keeps a second installation
@@ -145,25 +135,35 @@ fun SpawnScreen(
         val isCodex = current?.kind == "codex_cli"
         val isCodingCli = isClaude || isCodex
         val isLlama = current?.kind == "llama_cpp"
+        // Echo is the only kind with nothing to choose between.
+        val offersModels = isCodingCli || isLlama
+        // Where a session's tools act, which is the only thing a working directory decides.
+        val takesCwd = isCodingCli || isLlama
 
+        // Whichever machine and provider are chosen now, asked again when either changes. The
+        // previous answer is dropped first rather than left on screen: a model name from another
+        // machine looks exactly like one from this one.
         LaunchedEffect(machine?.id, current?.name) {
             model = ""
             providerModels = emptyList()
             providerModelsError = null
             permissionMode = current?.defaultPermissionMode.orEmpty()
-            if (isCodingCli) {
-                providerModelsLoading = true
-                try {
-                    providerModels =
-                        withContext(Dispatchers.IO) {
-                            fetchProviderModels(settings, machine.id, current.name)
-                        }
-                } catch (e: ApiException) {
-                    providerModelsError = e.message
-                } finally {
-                    providerModelsLoading = false
-                }
-            } else {
+            // Every kind that offers models at all, not only the coding CLIs: a llama provider
+            // answers with the GGUFs on the machine it runs on, through the same call. One
+            // question with one answer is what keeps the picker free of a branch on the kind.
+            if (machine == null || current == null || !offersModels) {
+                providerModelsLoading = false
+                return@LaunchedEffect
+            }
+            providerModelsLoading = true
+            try {
+                providerModels =
+                    withContext(Dispatchers.IO) {
+                        fetchProviderModels(settings, machine.id, current.name)
+                    }
+            } catch (e: ApiException) {
+                providerModelsError = e.message
+            } finally {
                 providerModelsLoading = false
             }
         }
@@ -220,29 +220,70 @@ fun SpawnScreen(
             modifier = Modifier.fillMaxWidth(),
         )
 
-        if (isLlama) {
-            // 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 on ${machine.name}. The Models screen downloads " +
-                        "to the backend; another machine needs the file put there itself.",
-                    style = MaterialTheme.typography.bodyMedium,
-                    color = MaterialTheme.colorScheme.onSurfaceVariant,
-                )
-            } else {
-                ChipGroup(
-                    label = "Model",
-                    // The file, not the whole key: the repository is the same for every
-                    // quantisation of a model, so the file name is what tells two of them apart.
-                    options = models.map { it.file },
-                    selected = models.firstOrNull { it.key == modelKey }?.file,
-                    onSelect = { file -> modelKey = models.first { it.file == file }.key },
-                )
+        if (offersModels) {
+            when {
+                providerModelsLoading ->
+                    Text(
+                        "Loading model choices…",
+                        style = MaterialTheme.typography.bodySmall,
+                        color = MaterialTheme.colorScheme.onSurfaceVariant,
+                    )
+                providerModelsError != null ->
+                    Text(
+                        "Model choices unavailable: $providerModelsError",
+                        style = MaterialTheme.typography.bodySmall,
+                        color = MaterialTheme.colorScheme.error,
+                    )
+                // A llama session cannot start without one, so this says what to do about it
+                // rather than only that there is nothing -- the models it needs are on the
+                // machine that will serve them, which is not always this backend.
+                providerModels.isEmpty() && isLlama ->
+                    Text(
+                        "No models on ${machine?.name}. The Models screen downloads " +
+                            "to the backend; another machine needs the file put there itself.",
+                        style = MaterialTheme.typography.bodyMedium,
+                        color = MaterialTheme.colorScheme.onSurfaceVariant,
+                    )
+                providerModels.isEmpty() ->
+                    Text(
+                        "This machine reported no selectable models.",
+                        style = MaterialTheme.typography.bodySmall,
+                        color = MaterialTheme.colorScheme.onSurfaceVariant,
+                    )
+                else -> {
+                    Spacer(Modifier.height(16.dp))
+                    ChipGroup(
+                        label = "Model",
+                        // The label, and the id is what is sent: for a llama model those differ,
+                        // since it is chosen by path and named by what is inside the file.
+                        options = providerModels.map { it.label },
+                        selected = providerModels.firstOrNull { it.id == model }?.label,
+                        onSelect = { chosen ->
+                            val id = providerModels.first { it.label == chosen }.id
+                            // A llama session has to have one, so choosing the same chip twice
+                            // must not clear it -- there is nothing to fall back to.
+                            model = if (model == id && !isLlama) "" else id
+                        },
+                    )
+                }
             }
             Spacer(Modifier.height(16.dp))
+        }
 
+        if (isCodingCli) {
+            // Free text as well as the chips above: the catalog is a shortcut, and a CLI will
+            // take a name it did not list.
+            OutlinedTextField(
+                value = model,
+                onValueChange = { model = it },
+                label = { Text("Model (blank = the CLI's default)") },
+                singleLine = true,
+                modifier = Modifier.fillMaxWidth(),
+            )
+            Spacer(Modifier.height(16.dp))
+        }
+
+        if (isLlama) {
             OutlinedTextField(
                 value = contextSize,
                 onValueChange = { contextSize = it },
@@ -260,48 +301,21 @@ fun SpawnScreen(
                 modifier = Modifier.fillMaxWidth(),
             )
             Spacer(Modifier.height(16.dp))
-        }
 
-        if (isCodingCli) {
-            when {
-                providerModelsLoading ->
-                    Text(
-                        "Loading model choices…",
-                        style = MaterialTheme.typography.bodySmall,
-                        color = MaterialTheme.colorScheme.onSurfaceVariant,
-                    )
-                providerModelsError != null ->
-                    Text(
-                        "Model choices unavailable: $providerModelsError",
-                        style = MaterialTheme.typography.bodySmall,
-                        color = MaterialTheme.colorScheme.error,
-                    )
-                providerModels.isEmpty() ->
-                    Text(
-                        "This machine reported no selectable models.",
-                        style = MaterialTheme.typography.bodySmall,
-                        color = MaterialTheme.colorScheme.onSurfaceVariant,
-                    )
-                else -> {
-                    Spacer(Modifier.height(16.dp))
-                    ChipGroup(
-                        label = "Model",
-                        options = providerModels,
-                        selected = model.ifEmpty { null },
-                        onSelect = { chosen -> model = if (model == chosen) "" else chosen },
-                    )
-                }
-            }
-            Spacer(Modifier.height(8.dp))
-            OutlinedTextField(
-                value = model,
-                onValueChange = { model = it },
-                label = { Text("Model (blank = the CLI's default)") },
-                singleLine = true,
-                modifier = Modifier.fillMaxWidth(),
+            // Said as what it is rather than as "MTP": the reader is choosing whether the session
+            // goes faster, and most models have nothing to turn on here at all.
+            ChipGroup(
+                label = "Speculative decoding (models that carry a draft head)",
+                options = listOf(SPECULATIVE_AUTO, SPECULATIVE_OFF),
+                selected = speculative,
+                onSelect = { speculative = it },
             )
             Spacer(Modifier.height(16.dp))
+        }
 
+        // Every session whose tools act on files needs one, which is both kinds that have
+        // tools -- a llama session's built-in tools run in it exactly as a CLI's do.
+        if (takesCwd) {
             OutlinedTextField(
                 value = cwd,
                 onValueChange = { cwd = it },
@@ -311,7 +325,11 @@ fun SpawnScreen(
                 modifier = Modifier.fillMaxWidth(),
             )
             Spacer(Modifier.height(16.dp))
+        }
 
+        // Offered wherever the provider has modes, rather than where this screen believes it
+        // does: the server is what knows, and llama.cpp grew them without this line changing.
+        if (current != null && current.permissionModes.isNotEmpty()) {
             ChipGroup(
                 label = "Permissions",
                 options = current.permissionModes,
@@ -319,7 +337,9 @@ fun SpawnScreen(
                 onSelect = { permissionMode = it },
             )
             Spacer(Modifier.height(16.dp))
+        }
 
+        if (isCodingCli) {
             // 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.
@@ -363,11 +383,9 @@ fun SpawnScreen(
                                     machine = machine.id,
                                     provider = chosen.name,
                                     title = title.trim(),
-                                    model =
-                                        if (isLlama) modelKey
-                                        else model.trim().takeIf { isCodingCli },
-                                    cwd = cwd.trim().takeIf { isCodingCli },
-                                    permissionMode = permissionMode.takeIf { isCodingCli },
+                                    model = model.trim().takeIf { offersModels },
+                                    cwd = cwd.trim().takeIf { takesCwd },
+                                    permissionMode = permissionMode.takeIf { it.isNotEmpty() },
                                     effort = effort.takeIf { isCodingCli },
                                     // Sent only when set, so blank means "whatever llama.cpp does
                                     // by default" rather than a zero.
@@ -382,6 +400,13 @@ fun SpawnScreen(
                                                     .trim()
                                                     .takeIf { it.isNotEmpty() }
                                                     ?.let { put("temperature", it) }
+                                                // Only the choice that changes anything: "auto"
+                                                // is the absence of the setting, not a value of
+                                                // it, so a session spawned without an opinion
+                                                // carries none.
+                                                if (speculative == SPECULATIVE_OFF) {
+                                                    put("speculative", "off")
+                                                }
                                             }
                                         },
                                 )
@@ -393,13 +418,23 @@ fun SpawnScreen(
                     }
                 }
             },
-            enabled = !busy && current != null && !(isLlama && modelKey == null),
+            // A llama session names the file to load, so there is nothing to spawn without one.
+            enabled = !busy && current != null && !(isLlama && model.isEmpty()),
         ) {
             Text(if (busy) "Spawning..." else "Spawn")
         }
     }
 }
 
+/**
+ * Leave the draft head to the server, which uses one wherever the model file has one. Spelled the
+ * same as the absence of the `speculative` parameter, because that is what it means.
+ */
+private const val SPECULATIVE_AUTO = "auto"
+
+/** The `speculative` parameter's only other value; see the llama driver's `SPECULATIVE`. */
+private const val SPECULATIVE_OFF = "off"
+
 /**
  * A labeled row of choices that wraps onto as many lines as it needs.
  *

server/src/config.rs

@@ -92,6 +92,32 @@ pub struct ProviderConfig {
     /// too; this is a shortcut list, not a restriction.
     #[serde(default, skip_serializing_if = "Vec::is_empty")]
     pub models: Vec<String>,
+    /// MCP servers whose tools this provider's sessions can use, on top of
+    /// whatever the provider runs itself.
+    ///
+    /// On the provider rather than the machine, because it is a statement
+    /// about what a session can do rather than about where it runs -- and
+    /// because only a driver that runs its own agent loop can use one. Today
+    /// that is llama.cpp; the coding CLIs have their own MCP configuration
+    /// and this would be a second, quieter answer to the same question.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub mcp_servers: Vec<McpServerConfig>,
+}
+
+/// An MCP server reached over HTTP.
+///
+/// A URL and nothing else: this backend connects to remote servers rather than
+/// spawning local ones, so there is no command, no arguments and no
+/// environment to configure. See `session::llama::mcp` for why that is the
+/// shape -- in short, it is what llama.cpp's own web UI does, and it keeps the
+/// tools on the machine with a route out rather than the machine with the GPU.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct McpServerConfig {
+    /// Prefixes every tool this server offers, so two servers with a `search`
+    /// are two tools. Also what a failure to connect is named by.
+    pub name: String,
+    pub url: String,
 }
 
 impl ProviderConfig {
@@ -278,7 +304,11 @@ impl DriverKind {
         match self {
             Self::ClaudeCli => &["manual", "acceptEdits", "auto", "bypassPermissions", "plan"],
             Self::CodexCli => &["workspace-write", "read-only", "danger-full-access"],
-            Self::Echo | Self::LlamaCpp => &[],
+            // Named by the driver that enforces them rather than repeated
+            // here: this list and the one the gate matches on being two
+            // literals is how a mode comes to be offered and then refused.
+            Self::LlamaCpp => crate::session::llama::PERMISSION_MODES,
+            Self::Echo => &[],
         }
     }
 
@@ -287,7 +317,8 @@ impl DriverKind {
         match self {
             Self::ClaudeCli => Some("auto"),
             Self::CodexCli => Some("workspace-write"),
-            Self::Echo | Self::LlamaCpp => None,
+            Self::LlamaCpp => Some(crate::session::llama::DEFAULT_PERMISSION_MODE),
+            Self::Echo => None,
         }
     }
 }
@@ -499,6 +530,7 @@ impl Config {
             kind: DriverKind::Echo,
             command: None,
             models: Vec::new(),
+            mcp_servers: Vec::new(),
         }
     }
 
@@ -579,6 +611,7 @@ mod tests {
                         kind: DriverKind::ClaudeCli,
                         command: Some("/usr/bin/claude".to_string()),
                         models: Vec::new(),
+                        mcp_servers: Vec::new(),
                     },
                 ]),
                 MachineConfig {
@@ -597,6 +630,7 @@ mod tests {
                         kind: DriverKind::ClaudeCli,
                         command: None,
                         models: vec!["haiku".to_string()],
+                        mcp_servers: Vec::new(),
                     }],
                 },
             ],
@@ -734,6 +768,7 @@ sessions: [(
                 kind: DriverKind::ClaudeCli,
                 command: Some("/usr/bin/claude".to_string()),
                 models: Vec::new(),
+                mcp_servers: Vec::new(),
             },
         ]);
         assert_eq!(

server/src/gguf.rs

@@ -0,0 +1,359 @@
+//! Just enough of the GGUF container to read a model's own name out of it.
+//!
+//! A `.gguf` file opens with a key/value table, and `general.name` in it is
+//! what the people who published the model called it -- "Qwen3-0.6B",
+//! "Qwen3.8-27B GSQ-RCO". Everything else this server knows a model by is
+//! filesystem trivia: `owner/repo/file.gguf` is where it was downloaded from,
+//! which is an address rather than a name, and on a phone it is a line of path
+//! where a word would do.
+//!
+//! **Read as far as the answer and no further.** The same table holds the
+//! tokenizer, which for a modern model is a 150,000-entry string array and
+//! most of several megabytes; `general.*` is written first by every converter
+//! in practice, so stopping at the name costs a few kilobytes instead. That is
+//! what makes this affordable to run over every model in a directory, and what
+//! lets the remote case work from a bounded prefix of the file rather than the
+//! whole of it.
+//!
+//! Anything unreadable is [`None`] rather than an error, at every level. A
+//! model with no name, a truncated prefix, a container version this does not
+//! know and a file that is not GGUF at all are one answer here -- "this file
+//! does not tell us" -- and the caller has a file name to fall back on. There
+//! is nothing a reader could do with the distinction.
+
+use std::io::Read;
+
+/// How many bytes of a model file are worth fetching to look for its name.
+///
+/// Only the remote path needs a number: a local read stops when it finds the
+/// key, but a file on another machine has to be asked for a fixed amount
+/// before anything can be parsed. Measured 2026-09-19 against the two models
+/// on this machine, `general.name` ends at byte **130** and **94** -- every
+/// converter writes `general.*` before the tokenizer arrays that make up the
+/// rest of the table. 8 KiB is two orders of magnitude of slack for that and
+/// still makes listing a directory of models one round trip's worth of bytes
+/// rather than a download, which is what decides the number: this is paid per
+/// model every time a spawn screen opens.
+pub const PREFIX_BYTES: u64 = 8 * 1024;
+
+/// The longest string this will allocate for, so a corrupt length field
+/// cannot ask for a gigabyte. Longer than any key or `general.*` value.
+const MAX_STRING: u64 = 64 * 1024;
+
+/// What `general.name` says, or `None` for every way of not finding out.
+///
+/// `read` is consumed only as far as the key: pass a file to read a local
+/// model, or a cursor over a prefix to read one whose bytes came from
+/// somewhere else.
+pub fn name(read: &mut impl Read) -> Option<String> {
+    match find(read, |key| key == "general.name") {
+        Some((STRING, read)) => string(read),
+        _ => None,
+    }
+}
+
+/// Whether this model carries a multi-token-prediction head.
+///
+/// Worth asking because `llama-server` **exits** when told to use one that is
+/// not there -- `--spec-type draft-mtp` on a plain model is "context type MTP
+/// requested but model doesn't contain MTP layers" and then a server that
+/// never comes up. So the flag can only be passed once this has said yes, and
+/// a `false` here is the same answer as an unreadable file: don't ask for it.
+///
+/// Matched on the key's tail rather than its whole name, because the key is
+/// prefixed with the architecture (`qwen35.nextn_predict_layers`) and the
+/// architecture is whatever the next model is. The value is not read: a model
+/// that declares the key at all is one whose tensors carry the head, and the
+/// two disagreeing is a broken file rather than a state to handle.
+pub fn has_mtp_head(read: &mut impl Read) -> bool {
+    find(read, |key| key.ends_with(".nextn_predict_layers")).is_some()
+}
+
+/// Steps through the metadata table to the first key `wanted` accepts,
+/// returning its value's type tag and the reader positioned at the value.
+fn find<R: Read>(read: &mut R, wanted: impl Fn(&str) -> bool) -> Option<(u32, &mut R)> {
+    let mut magic = [0u8; 4];
+    read.read_exact(&mut magic).ok()?;
+    if &magic != b"GGUF" {
+        return None;
+    }
+    let _version = u32s(read)?;
+    let _tensors = u64s(read)?;
+    let count = u64s(read)?;
+    for _ in 0..count {
+        let found = string(read)?;
+        let kind = u32s(read)?;
+        if wanted(&found) {
+            return Some((kind, read));
+        }
+        skip_value(kind, read)?;
+    }
+    None
+}
+
+// The value type tags, in the container's own numbering. Only the two this
+// has to act on are named; the rest are widths, and `scalar_width` is where
+// the numbering is written down once.
+const STRING: u32 = 8;
+const ARRAY: u32 = 9;
+
+/// How many bytes a scalar of this type occupies, or `None` for a type that
+/// is not a scalar -- which includes a tag this build does not know, since a
+/// value of unknown length cannot be stepped over.
+fn scalar_width(kind: u32) -> Option<u64> {
+    match kind {
+        // u8, i8, bool
+        0 | 1 | 7 => Some(1),
+        // u16, i16
+        2 | 3 => Some(2),
+        // u32, i32, f32
+        4..=6 => Some(4),
+        // u64, i64, f64
+        10..=12 => Some(8),
+        _ => None,
+    }
+}
+
+/// Steps over one value of `kind` without keeping it.
+///
+/// Recursive only in the sense that an array's elements are values; GGUF
+/// arrays do not nest, so the recursion is one level deep by construction.
+fn skip_value(kind: u32, read: &mut impl Read) -> Option<()> {
+    match kind {
+        STRING => {
+            let len = u64s(read)?;
+            skip(len, read)
+        }
+        ARRAY => {
+            let element = u32s(read)?;
+            let count = u64s(read)?;
+            match scalar_width(element) {
+                // The whole array at once: this is the tokenizer's scores and
+                // token types, and stepping over them one at a time is a
+                // syscall per token.
+                Some(width) => skip(count.checked_mul(width)?, read),
+                None if element == STRING => {
+                    for _ in 0..count {
+                        let len = u64s(read)?;
+                        skip(len, read)?;
+                    }
+                    Some(())
+                }
+                // An array of arrays, or of something this build has no width
+                // for: the rest of the table can no longer be located.
+                None => None,
+            }
+        }
+        _ => skip(scalar_width(kind)?, read),
+    }
+}
+
+/// Discards `count` bytes, failing if the input ends first.
+///
+/// Chunked against a bounded buffer rather than read into a `Vec` of the
+/// stated size: the sizes here come out of the file, and the file may be a
+/// truncated prefix or not a GGUF at all.
+fn skip(count: u64, read: &mut impl Read) -> Option<()> {
+    let mut scratch = [0u8; 8192];
+    let mut left = count;
+    while left > 0 {
+        let want = left.min(scratch.len() as u64) as usize;
+        read.read_exact(&mut scratch[..want]).ok()?;
+        left -= want as u64;
+    }
+    Some(())
+}
+
+fn string(read: &mut impl Read) -> Option<String> {
+    let len = u64s(read)?;
+    if len > MAX_STRING {
+        return None;
+    }
+    let mut bytes = vec![0u8; len as usize];
+    read.read_exact(&mut bytes).ok()?;
+    String::from_utf8(bytes).ok()
+}
+
+fn u32s(read: &mut impl Read) -> Option<u32> {
+    let mut bytes = [0u8; 4];
+    read.read_exact(&mut bytes).ok()?;
+    Some(u32::from_le_bytes(bytes))
+}
+
+fn u64s(read: &mut impl Read) -> Option<u64> {
+    let mut bytes = [0u8; 8];
+    read.read_exact(&mut bytes).ok()?;
+    Some(u64::from_le_bytes(bytes))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Builds a GGUF header holding exactly these keys, so the parser is
+    /// tested against the layout rather than against a fixture nobody here
+    /// can regenerate.
+    fn header(entries: &[(&str, Value)]) -> Vec<u8> {
+        let mut out = Vec::from(*b"GGUF");
+        out.extend(3u32.to_le_bytes());
+        out.extend(0u64.to_le_bytes());
+        out.extend((entries.len() as u64).to_le_bytes());
+        for (key, value) in entries {
+            put_string(&mut out, key);
+            value.write(&mut out);
+        }
+        out
+    }
+
+    enum Value {
+        Str(&'static str),
+        U32(u32),
+        Strings(Vec<&'static str>),
+        Floats(Vec<f32>),
+    }
+
+    impl Value {
+        fn write(&self, out: &mut Vec<u8>) {
+            match self {
+                Self::Str(text) => {
+                    out.extend(STRING.to_le_bytes());
+                    put_string(out, text);
+                }
+                Self::U32(number) => {
+                    out.extend(4u32.to_le_bytes());
+                    out.extend(number.to_le_bytes());
+                }
+                Self::Strings(items) => {
+                    out.extend(ARRAY.to_le_bytes());
+                    out.extend(STRING.to_le_bytes());
+                    out.extend((items.len() as u64).to_le_bytes());
+                    for item in items {
+                        put_string(out, item);
+                    }
+                }
+                Self::Floats(items) => {
+                    out.extend(ARRAY.to_le_bytes());
+                    out.extend(6u32.to_le_bytes());
+                    out.extend((items.len() as u64).to_le_bytes());
+                    for item in items {
+                        out.extend(item.to_le_bytes());
+                    }
+                }
+            }
+        }
+    }
+
+    fn put_string(out: &mut Vec<u8>, text: &str) {
+        out.extend((text.len() as u64).to_le_bytes());
+        out.extend(text.as_bytes());
+    }
+
+    #[test]
+    fn the_name_is_read_past_every_other_kind_of_value() {
+        let bytes = header(&[
+            ("general.architecture", Value::Str("qwen3")),
+            ("general.file_type", Value::U32(7)),
+            ("qwen3.attention.head_count", Value::U32(16)),
+            ("tokenizer.ggml.scores", Value::Floats(vec![0.5; 64])),
+            ("tokenizer.ggml.tokens", Value::Strings(vec!["a", "b", "c"])),
+            ("general.name", Value::Str("Qwen3-0.6B")),
+        ]);
+        assert_eq!(
+            name(&mut bytes.as_slice()),
+            Some("Qwen3-0.6B".to_string()),
+            "every value before the name has to be steppable over",
+        );
+    }
+
+    #[test]
+    /// The remote case: a prefix is all there is, and running off the end of
+    /// it is "we don't know" rather than a failure worth reporting. The
+    /// caller has the file name.
+    fn a_truncated_file_has_no_name_rather_than_failing() {
+        let bytes = header(&[
+            ("tokenizer.ggml.tokens", Value::Strings(vec!["a", "b", "c"])),
+            ("general.name", Value::Str("Qwen3-0.6B")),
+        ]);
+        for cut in [4, 12, 24, bytes.len() - 4] {
+            assert_eq!(name(&mut &bytes[..cut]), None, "cut at {cut}");
+        }
+    }
+
+    #[test]
+    fn a_file_that_is_not_gguf_has_no_name() {
+        assert_eq!(name(&mut b"not a model at all".as_slice()), None);
+        assert_eq!(name(&mut b"".as_slice()), None);
+    }
+
+    #[test]
+    /// A name that is not a string is not a name. The alternative is
+    /// rendering a number as one, which reads as a model called "7".
+    fn a_name_of_the_wrong_type_is_not_read() {
+        let bytes = header(&[("general.name", Value::U32(7))]);
+        assert_eq!(name(&mut bytes.as_slice()), None);
+    }
+
+    #[test]
+    /// The head is found by the tail of the key, because the whole key is
+    /// prefixed with whatever architecture the model is.
+    fn an_mtp_head_is_found_whatever_the_architecture_is_called() {
+        let with = header(&[
+            ("general.architecture", Value::Str("qwen35")),
+            ("qwen35.block_count", Value::U32(64)),
+            ("qwen35.nextn_predict_layers", Value::U32(1)),
+        ]);
+        assert!(has_mtp_head(&mut with.as_slice()));
+
+        let without = header(&[
+            ("general.architecture", Value::Str("qwen3")),
+            ("qwen3.block_count", Value::U32(28)),
+        ]);
+        assert!(!has_mtp_head(&mut without.as_slice()));
+    }
+
+    #[test]
+    /// A prefix that stops short says no, and that is the direction it has to
+    /// fail in: `--spec-type draft-mtp` on a model with no head is a server
+    /// that exits, so "we could not tell" and "it has none" both mean don't
+    /// ask for it.
+    fn a_truncated_file_reports_no_mtp_head() {
+        let bytes = header(&[("qwen35.nextn_predict_layers", Value::U32(1))]);
+        assert!(!has_mtp_head(&mut &bytes[..12]));
+    }
+
+    #[test]
+    /// The real thing, when this machine happens to have one. Skipped rather
+    /// than failed where it does not: the models directory is not part of the
+    /// checkout, and a test that needs gigabytes to run is one nobody runs.
+    fn a_real_model_on_this_machine_reads_back_its_name() {
+        let Some(home) = std::env::var_os("HOME") else {
+            return;
+        };
+        let dir = std::path::Path::new(&home).join(".local/share/ai-app/models");
+        let mut found = Vec::new();
+        collect_gguf(&dir, &mut found);
+        for path in found {
+            let mut file = std::fs::File::open(&path).expect("open");
+            let read = name(&mut file);
+            assert!(
+                read.is_some_and(|name| !name.trim().is_empty()),
+                "{} has a name in it and this did not read one",
+                path.display(),
+            );
+        }
+    }
+
+    fn collect_gguf(dir: &std::path::Path, found: &mut Vec<std::path::PathBuf>) {
+        let Ok(entries) = std::fs::read_dir(dir) else {
+            return;
+        };
+        for entry in entries.flatten() {
+            let path = entry.path();
+            if path.is_dir() {
+                collect_gguf(&path, found);
+            } else if path.extension().is_some_and(|e| e == "gguf") {
+                found.push(path);
+            }
+        }
+    }
+}

server/src/machines.rs

@@ -15,6 +15,7 @@
 //! phone is not being given.
 
 use anyhow::{Context, Result};
+use serde::Serialize;
 use serde_json::{Value, json};
 
 use crate::config::{DriverKind, ProviderConfig};
@@ -57,12 +58,7 @@ pub async fn discover(transport: &Transport) -> Result<Vec<ProviderConfig>> {
     // and nowhere else. Offering it on a remote machine would be a choice that
     // changes nothing.
     if matches!(transport, Transport::Here) {
-        providers.push(ProviderConfig {
-            name: crate::config::ECHO_PROVIDER.to_string(),
-            kind: DriverKind::Echo,
-            command: None,
-            models: Vec::new(),
-        });
+        providers.push(crate::config::Config::echo_provider());
     }
     for (name, binary, kind) in PROBES {
         let path = found
@@ -83,22 +79,88 @@ pub async fn discover(transport: &Transport) -> Result<Vec<ProviderConfig>> {
                 DriverKind::ClaudeCli => CLAUDE_MODELS.iter().map(|m| (*m).to_string()).collect(),
                 _ => Vec::new(),
             },
+            mcp_servers: mcp_defaults(*kind),
         });
     }
     Ok(providers)
 }
 
+/// One model a picker can offer, and what to call it there.
+///
+/// Two fields rather than one string because for one provider they differ:
+/// a llama.cpp model is chosen by the path it lives at and read as the name
+/// its own metadata gives it. Every other provider's id is already the name,
+/// and says so by repeating it -- which is what keeps the picker free of a
+/// branch on the session kind.
+#[derive(Debug, Clone, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct OfferedModel {
+    /// What a spawn or a model change is given. Opaque to the phone.
+    pub id: String,
+    /// What a person reads on the chip.
+    pub label: String,
+}
+
+impl OfferedModel {
+    /// A model whose id is its own name, which is every provider but llama.
+    fn plain(id: impl Into<String>) -> Self {
+        let id = id.into();
+        Self {
+            label: id.clone(),
+            id,
+        }
+    }
+}
+
+/// The MCP servers a newly discovered provider of this kind starts with.
+///
+/// A default rather than something to be typed in: a llama session with no web
+/// search is the state somebody would then have to find out how to leave, and
+/// Exa is what llama.cpp's own web UI offers under the same name. It is an
+/// ordinary config entry once written, so removing it is deleting a line.
+///
+/// Only llama.cpp, because only a driver that runs its own agent loop can use
+/// one -- the coding CLIs configure MCP themselves and a second answer here
+/// would quietly disagree with theirs.
+fn mcp_defaults(kind: DriverKind) -> Vec<crate::config::McpServerConfig> {
+    match kind {
+        DriverKind::LlamaCpp => vec![crate::config::McpServerConfig {
+            name: "exa".to_string(),
+            url: crate::session::llama::EXA_MCP_URL.to_string(),
+        }],
+        _ => Vec::new(),
+    }
+}
+
 /// Models the selected provider currently offers on this machine.
 ///
 /// Codex's catalog is account- and CLI-version-specific, so it is asked at the
-/// moment the picker opens rather than copied into `config.ron`. Other
-/// providers retain the shortcut list discovery stored for them.
+/// moment the picker opens rather than copied into `config.ron`. A llama.cpp
+/// provider offers the GGUFs on the machine it runs on, through this same call
+/// -- there was a second route answering that alone, and it went when this one
+/// learned to, because a picker offering a model the spawn screen does not, or
+/// naming it differently, is two answers to one question. Other providers
+/// retain the shortcut list discovery stored for them.
 pub async fn provider_models(
     transport: &Transport,
     provider: &ProviderConfig,
-) -> Result<Vec<String>> {
+    models_dir: &std::path::Path,
+) -> Result<Vec<OfferedModel>> {
+    if provider.kind == DriverKind::LlamaCpp {
+        let dir = crate::models::dir_on(transport, models_dir);
+        let found = crate::models::on_machine(transport, &dir).await?;
+        let labels = crate::models::labels(&found);
+        return Ok(found
+            .into_iter()
+            .zip(labels)
+            .map(|(model, label)| OfferedModel {
+                id: model.key,
+                label,
+            })
+            .collect());
+    }
     if provider.kind != DriverKind::CodexCli {
-        return Ok(provider.models.clone());
+        return Ok(provider.models.iter().map(OfferedModel::plain).collect());
     }
     let transport = transport.clone();
     let program = provider.program().to_string();
@@ -114,7 +176,10 @@ pub async fn provider_models(
             json!({"id": 2, "method": "model/list", "params": {"includeHidden": false, "limit": 100}}),
         ];
         let answer = transport.request_json_blocking(&launch, &initial, &requests, 2)?;
-        parse_codex_models(&answer)
+        Ok(parse_codex_models(&answer)?
+            .into_iter()
+            .map(OfferedModel::plain)
+            .collect())
     })
     .await?
 }

server/src/main.rs

@@ -14,6 +14,7 @@
 mod auth;
 mod config;
 mod files;
+mod gguf;
 mod machines;
 mod media;
 mod models;

server/src/models.rs

@@ -50,6 +50,71 @@ pub struct LocalModel {
     pub repo: String,
     pub file: String,
     pub bytes: u64,
+    /// What the file says it is called (`general.name` in its own metadata),
+    /// absent when it does not say or could not be read. Not a label: see
+    /// [`labels`] for what a reader is actually shown, which needs the rest of
+    /// the list to decide.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub name: Option<String>,
+}
+
+/// What each of these models should be called on screen, in the same order.
+///
+/// A model's own name is the best answer and is not always an answer at all:
+/// two quantisations of one model carry the same `general.name`, and a chip
+/// row with two identical chips is one you cannot choose from. So this is a
+/// cascade -- the model's own name, else its file name, else its full key --
+/// and each model takes the first rung that nothing else on this machine
+/// shares. The last rung always terminates it, because the key is what makes
+/// these unique in the first place.
+///
+/// Decided over the whole list rather than per model because ambiguity is a
+/// property of the set: the same file is unambiguous on a machine holding one
+/// quantisation and not on a machine holding three, and only the list knows
+/// which machine this is.
+pub fn labels(models: &[LocalModel]) -> Vec<String> {
+    let rungs = |model: &LocalModel| {
+        [
+            model.name.clone(),
+            Some(model.file.trim_end_matches(".gguf").to_string()),
+            Some(model.key.clone()),
+        ]
+    };
+    let mut taken: Vec<HashMap<String, usize>> = vec![HashMap::new(); 3];
+    for model in models {
+        for (rung, candidate) in rungs(model).into_iter().enumerate() {
+            if let Some(candidate) = candidate {
+                *taken[rung].entry(candidate).or_insert(0) += 1;
+            }
+        }
+    }
+    models
+        .iter()
+        .map(|model| {
+            rungs(model)
+                .into_iter()
+                .enumerate()
+                .find_map(|(rung, candidate)| {
+                    let candidate = candidate?;
+                    (taken[rung].get(&candidate) == Some(&1)).then_some(candidate)
+                })
+                // Unreachable: the key rung is unique by construction. Said as
+                // the key rather than as a panic, because a duplicate key would
+                // mean the same file listed twice and a name is still the
+                // honest thing to draw for it.
+                .unwrap_or_else(|| model.key.clone())
+        })
+        .collect()
+}
+
+/// The model's own name, read out of the file itself.
+///
+/// Absent for every way of not finding out -- see [`crate::gguf`]. The file is
+/// opened and read only as far as the name, which is the first few hundred
+/// bytes, so this is affordable once per model per listing.
+fn name_of(path: &Path) -> Option<String> {
+    let mut file = std::fs::File::open(path).ok()?;
+    crate::gguf::name(&mut file)
 }
 
 /// What a run is doing, or did. Flat rather than a tagged enum carrying its
@@ -519,6 +584,7 @@ fn collect(root: &Path, dir: &Path, found: &mut Vec<LocalModel>) {
             repo: repo.to_string(),
             file: file.to_string(),
             bytes: entry.metadata().map(|m| m.len()).unwrap_or(0),
+            name: name_of(&path),
         });
     }
 }
@@ -566,33 +632,46 @@ pub fn dir_on(transport: &Transport, local: &Path) -> String {
 /// 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.
+///
+/// Each record carries the head of the file as well as its size, because a
+/// model's own name is inside it (see [`crate::gguf`]) and the file is on the
+/// far machine. The alternative is a second round trip per model, or naming
+/// remote models by path while local ones get their proper names -- one
+/// machine's models reading differently from another's is exactly the
+/// confusion the name was added to remove. The prefix is bounded at
+/// [`crate::gguf::PREFIX_BYTES`], which is what keeps this one round trip's
+/// worth of bytes.
 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 script = format!(
+        "p=$1; case $p in \"~\") p=$HOME;; \"~/\"*) p=$HOME/${{p#\"~/\"}};; esac; \
+         [ -d \"$p\" ] || exit 0; cd \"$p\" || exit 0; \
+         find . -type f -name '*.gguf' -exec sh -c '\
+           for f do printf \"%s\\t%s\\t%s\\0\" \"$(wc -c < \"$f\")\" \
+             \"$(head -c {prefix} \"$f\" | base64 | tr -d \"\\n\")\" \"${{f#./}}\"; done\
+         ' sh {{}} +",
+        prefix = crate::gguf::PREFIX_BYTES,
+    );
     let launch = Launch::new(
         "sh",
-        vec![
-            "-c".to_string(),
-            script.to_string(),
-            "sh".to_string(),
-            dir.to_string(),
-        ],
+        vec!["-c".to_string(), script, "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)| {
+        // Three fields, and the name last, so a `\t` in a filename survives.
+        // The middle one is base64, which has no tab in its alphabet.
+        .filter_map(|record| {
+            let (bytes, rest) = record.split_once('\t')?;
+            let (head, key) = rest.split_once('\t')?;
             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),
+                name: name_in_prefix(head),
             })
         })
         .collect();
@@ -600,6 +679,18 @@ pub async fn on_machine(transport: &Transport, dir: &str) -> Result<Vec<LocalMod
     Ok(found)
 }
 
+/// The model's name out of a base64 prefix of its file.
+///
+/// The remote half of [`name_of`], and `None` for everything that half
+/// answers `None` for, plus a prefix that did not survive the trip.
+fn name_in_prefix(head: &str) -> Option<String> {
+    use base64::Engine as _;
+    let bytes = base64::engine::general_purpose::STANDARD
+        .decode(head.trim())
+        .ok()?;
+    crate::gguf::name(&mut bytes.as_slice())
+}
+
 /// A model repository on HuggingFace, as the browse screen shows it.
 #[derive(Debug, Clone, Serialize)]
 #[serde(rename_all = "camelCase")]

server/src/provider_auth.rs

@@ -436,6 +436,7 @@ mod tests {
             kind: DriverKind::ClaudeCli,
             command: Some(cli.display().to_string()),
             models: Vec::new(),
+            mcp_servers: Vec::new(),
         };
 
         let started = logins.start(machine, provider);

server/src/routes.rs

@@ -7,8 +7,7 @@
 //! POST   /machines                       add {name, ssh?} -- providers are discovered
 //! POST   /machines/probe                 dry run {ssh?}: what would be found there
 //! GET    /machines/{id}                  one machine, for refetching after a change
-//! GET    /machines/{id}/models           GGUFs on that machine, for a llama session
-//! GET    /machines/{id}/providers/{provider}/models  models a CLI currently offers
+//! GET    /machines/{id}/providers/{provider}/models  models that provider offers
 //! POST   /machines/{id}/providers/{provider}/auth    begin provider sign-in
 //! GET    /machines/{id}/providers/{provider}/auth/{attempt}  sign-in state
 //! POST   /machines/{id}/providers/{provider}/auth/{attempt}/code  submit browser code
@@ -131,7 +130,6 @@ pub fn router(manager: Arc<SessionManager>) -> Router {
             get(read_machine).put(update_machine).delete(delete_machine),
         )
         // The models on a configured machine, for a llama session there.
-        .route("/machines/{id}/models", get(machine_models))
         .route(
             "/machines/{id}/providers/{provider}/models",
             get(provider_models),
@@ -565,40 +563,19 @@ 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 machine 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 machine_models(
-    State(manager): State<Arc<SessionManager>>,
-    UrlPath(id): UrlPath<String>,
-) -> Result<axum::Json<Vec<crate::models::LocalModel>>, ApiError> {
-    let machine = machine_by_id(&manager, &id)?;
-    let transport = crate::session::transport::Transport::for_machine(&machine);
-    let dir = crate::models::dir_on(&transport, manager.models_dir());
-    crate::models::on_machine(&transport, &dir)
-        .await
-        .map(axum::Json)
-        .map_err(from_machine)
-}
-
-/// The models a CLI provider currently offers on its configured machine.
-/// Codex answers from its live account catalog; providers with a configured
-/// shortcut list return that list.
+/// The models a provider currently offers on its configured machine.
+/// Codex answers from its live account catalog, llama.cpp from the GGUFs on
+/// that machine, and providers with a configured shortcut list return it.
 async fn provider_models(
     State(manager): State<Arc<SessionManager>>,
     UrlPath((id, provider_name)): UrlPath<(String, String)>,
-) -> Result<axum::Json<Vec<String>>, ApiError> {
+) -> Result<axum::Json<Vec<crate::machines::OfferedModel>>, ApiError> {
     let machine = machine_by_id(&manager, &id)?;
     let provider = machine.provider(&provider_name).ok_or_else(|| {
         ApiError::NotFound(format!("no provider {provider_name} on {}", machine.name))
     })?;
     let transport = crate::session::transport::Transport::for_machine(&machine);
-    crate::machines::provider_models(&transport, provider)
+    crate::machines::provider_models(&transport, provider, manager.models_dir())
         .await
         .map(axum::Json)
         .map_err(from_machine)

server/src/session/driver.rs

@@ -564,6 +564,22 @@ pub enum SessionStatus {
     Running,
     AwaitingInput,
     Compacting,
+    /// The session's process is up but cannot be spoken to yet.
+    ///
+    /// Its own state because the two it would otherwise borrow are both
+    /// wrong in ways somebody notices. `Running` means the session is
+    /// answering, so a model taking a minute to load looks like a model
+    /// thinking for a minute -- and there is no way to tell from the screen
+    /// that the first message will be refused. `Idle` invites that message
+    /// and then loses it.
+    ///
+    /// It exists for `llama-server`, which reads a multi-gigabyte file off
+    /// disk before it answers anything, and it is general because the
+    /// condition is: a process that is started and not yet ready is a state
+    /// any driver may have to report. Nothing is queued *because* of this
+    /// state -- a driver that reports it is responsible for holding what it
+    /// is sent until it can deliver it -- but this is what says so on screen.
+    Loading,
     /// The session's own turn is over, but work it started is still going:
     /// a backgrounded subagent, or a command left running.
     ///

server/src/session/llama.rs

@@ -1,907 +0,0 @@
-//! The llama.cpp driver: a `llama-server` process per session, spoken to over
-//! its OpenAI-compatible HTTP API and translated into the common event model.
-//!
-//! Two things make this shaped differently from the Claude driver.
-//!
-//! **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. 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 machine
-//! 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
-//! in this struct, which is not tidiness: a copy in driver memory is invisible
-//! to a second device and gone when this process restarts.
-//!
-//! That leaves the Claude driver as the odd one out rather than this one -- the
-//! CLI's own memory of a conversation is a cache in front of the same
-//! transcript. Resolve any inconsistency in this direction.
-
-use std::path::{Path, PathBuf};
-use std::sync::Arc;
-use std::sync::atomic::{AtomicBool, Ordering};
-
-use anyhow::{Context, Result, bail};
-use serde::{Deserialize, Serialize};
-use serde_json::json;
-
-use super::driver::{AttachmentRef, Driver, Event, EventSink, SessionStatus};
-use super::process;
-use super::transport::{Launch, Streams, Transport};
-use crate::config::{ProviderConfig, SessionConfig};
-
-/// How long to wait for a model to load before giving up. Loading is mostly
-/// disk, and a large quantised model on a cold cache is genuinely slow, so this
-/// is generous -- the failure it exists for is a server that will never answer.
-const READY_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(300);
-
-/// One turn in the conversation this driver keeps on the server's behalf.
-#[derive(Debug, Clone, Serialize, Deserialize)]
-struct Message {
-    role: String,
-    content: String,
-}
-
-pub struct LlamaDriver {
-    sink: EventSink,
-    /// Where this session's own llama-server answers.
-    endpoint: String,
-    /// Where the conversation is read back from, one line per event.
-    transcript: PathBuf,
-    /// Sampling settings chosen at spawn, sent with every request.
-    sampling: serde_json::Map<String, serde_json::Value>,
-    /// Set by [`Driver::interrupt`]; the streaming loop checks it between
-    /// chunks and stops, leaving what was generated in the transcript.
-    cancel: Arc<AtomicBool>,
-    /// Where this session's process record lives, so [`Driver::stop`] can find
-    /// the server it has to end.
-    session_dir: PathBuf,
-}
-
-impl LlamaDriver {
-    /// Takes charge of this session's `llama-server`: the one already loaded if
-    /// there is one, otherwise a new one.
-    ///
-    /// One entry point, for the reason `ClaudeDriver::launch` gives, expensive
-    /// 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,
-        transport: &Transport,
-        models_dir: &Path,
-        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> {
-        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_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
-        // has wedged still reports as a failure rather than as a session that
-        // silently never replies.
-        if let Some(process::Record {
-            detail: process::Detail::Http { port },
-            pid,
-            ..
-        }) = process::live(session_dir)
-        {
-            tracing::info!(
-                "session {} reattaching to the llama-server it left loaded (pid {pid}, port {port})",
-                meta.id
-            );
-            return Ok(Self::attached(
-                format!("http://127.0.0.1:{port}"),
-                meta,
-                model,
-                transcript,
-                session_dir,
-                sink,
-            ));
-        }
-
-        // 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.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(),
-            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
-        // them later needn't reload anything.
-        for (key, flag) in [
-            ("contextSize", "-c"),
-            ("gpuLayers", "-ngl"),
-            ("threads", "-t"),
-        ] {
-            if let Some(value) = meta.params.get(key) {
-                args.push(flag.to_string());
-                args.push(value.clone());
-            }
-        }
-
-        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
-        // of why.
-        let child = transport.spawn(
-            &launch,
-            Streams::Detached {
-                stdin: std::process::Stdio::null(),
-                stdout: log_file(&session_dir.join(SERVER_LOG))?.into(),
-                stderr: log_file(&session_dir.join(SERVER_LOG))?.into(),
-            },
-        )?;
-        let pid = child
-            .id()
-            .context("llama-server exited before it could be recorded")?;
-        tracing::info!(
-            "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
-        // session is alive, because after a restart there is no `Child` to ask.
-        tokio::spawn(async move {
-            let mut child = child;
-            let _ = child.wait().await;
-        });
-
-        // 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:{}", forward.here),
-            meta,
-            model,
-            transcript,
-            session_dir,
-            sink,
-        ))
-    }
-
-    /// The driver for a `llama-server` at `endpoint`, however it got there.
-    ///
-    /// Shared by starting one and adopting one, because everything after "there
-    /// is a server at this address" is identical -- including waiting for it to
-    /// answer, which an adopted one still owes: a recorded pid says a process
-    /// exists, not that its model is loaded.
-    fn attached(
-        endpoint: String,
-        meta: &SessionConfig,
-        model: &str,
-        transcript: &Path,
-        session_dir: &Path,
-        sink: EventSink,
-    ) -> Self {
-        // Loading is slow enough to be worth saying so: the session shows as
-        // running until the model is in memory, rather than looking ready and
-        // refusing the first message.
-        let _ = sink.send(Event::Status {
-            state: SessionStatus::Running,
-        });
-        {
-            let sink = sink.clone();
-            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, &session_dir) {
-                Ok(()) => {
-                    tracing::info!("{model} loaded and answering at {endpoint}");
-                    let _ = sink.send(Event::Status {
-                        state: SessionStatus::Idle,
-                    });
-                    watch(session_dir, sink);
-                }
-                Err(err) => {
-                    let _ = sink.send(Event::Error {
-                        message: format!("{model} never became ready: {err:#}"),
-                    });
-                    let _ = sink.send(Event::Status {
-                        state: SessionStatus::Exited,
-                    });
-                    process::clear(&session_dir);
-                }
-            });
-        }
-
-        let mut sampling = serde_json::Map::new();
-        for (key, field) in [
-            ("temperature", "temperature"),
-            ("topP", "top_p"),
-            ("topK", "top_k"),
-            ("maxTokens", "max_tokens"),
-        ] {
-            if let Some(raw) = meta.params.get(key)
-                && let Ok(number) = raw.parse::<f64>()
-            {
-                sampling.insert(field.to_string(), json!(number));
-            }
-        }
-
-        Self {
-            sink,
-            endpoint,
-            transcript: transcript.to_path_buf(),
-            sampling,
-            cancel: Arc::new(AtomicBool::new(false)),
-            session_dir: session_dir.to_path_buf(),
-        }
-    }
-}
-
-/// Where llama-server's own output goes. One file for both streams: it is
-/// diagnostics nobody parses, and interleaving them is how it reads in a
-/// terminal anyway.
-const SERVER_LOG: &str = "llama-server.log";
-
-/// How often a loaded server is checked for still being there. Slower than the
-/// Claude driver's stdout poll because nothing is waiting on it: this only has
-/// to notice a server that has gone.
-const WATCH_INTERVAL: std::time::Duration = std::time::Duration::from_secs(2);
-
-/// An owner-only log opened for appending, so the two streams pointed at
-/// it do not overwrite each other and a reattach keeps what came before.
-fn log_file(path: &Path) -> Result<std::fs::File> {
-    use std::os::unix::fs::OpenOptionsExt;
-    std::fs::OpenOptions::new()
-        .create(true)
-        .append(true)
-        .mode(0o600)
-        .open(path)
-        .with_context(|| format!("opening {}", path.display()))
-}
-
-/// Reports the server going away, for as long as the session is there to report
-/// it to.
-///
-/// Polled rather than waited on, for the reason the Claude driver gives: after a
-/// restart this server is not the process's parent, so liveness has to be a
-/// question asked of the record -- and asking it two different ways is how the
-/// two answers come to disagree.
-fn watch(session_dir: PathBuf, sink: EventSink) {
-    std::thread::spawn(move || {
-        loop {
-            std::thread::sleep(WATCH_INTERVAL);
-            match process::recorded(&session_dir) {
-                Some((_, process::Liveness::Alive)) => {}
-                // Nothing recorded means the session was stopped or deleted
-                // deliberately, and whoever did that has already said so.
-                None => return,
-                Some((_, process::Liveness::Dead)) => {
-                    if !process::stopping(&session_dir) {
-                        let _ = sink.send(Event::Error {
-                            message: "llama-server exited".to_string(),
-                        });
-                    }
-                    let _ = sink.send(Event::Status {
-                        state: SessionStatus::Exited,
-                    });
-                    process::clear(&session_dir);
-                    return;
-                }
-                Some((_, process::Liveness::Unknown)) => {
-                    let _ = sink.send(Event::Status {
-                        state: SessionStatus::Unknown,
-                    });
-                }
-            }
-            if sink.is_closed() {
-                return;
-            }
-        }
-    });
-}
-
-impl Driver for LlamaDriver {
-    fn send_user_message(&self, text: String, attachments: Vec<AttachmentRef>) {
-        if !attachments.is_empty() {
-            let _ = self.sink.send(Event::Error {
-                message: "this model can't be sent attachments or files".to_string(),
-            });
-        }
-        let sink = self.sink.clone();
-        let endpoint = self.endpoint.clone();
-        let transcript = self.transcript.clone();
-        let sampling = self.sampling.clone();
-        let cancel = Arc::clone(&self.cancel);
-        cancel.store(false, Ordering::Relaxed);
-
-        // Its own thread: the request blocks for as long as the model takes to
-        // generate, which is the whole point of streaming it.
-        std::thread::spawn(move || {
-            // Nothing is ever held back here -- there is no queue to wait in --
-            // so the message is taken the moment it arrives. Said anyway,
-            // because this is what records it: see `MessageTaken`.
-            let _ = sink.send(Event::MessageTaken {
-                id: None,
-                text: text.clone(),
-                // Never any: this driver refuses attachments above.
-                attachments: Vec::new(),
-            });
-            let _ = sink.send(Event::Status {
-                state: SessionStatus::Running,
-            });
-            // Everything before this message, plus this message. Read rather
-            // than remembered, and `text` is appended here rather than waited
-            // for, because the message's own transcript entry is still on its
-            // way when this runs.
-            let mut messages = conversation(&transcript);
-            messages.push(Message {
-                role: "user".into(),
-                content: text,
-            });
-            // The reply is not stored: the deltas below are the durable record,
-            // so the next turn reads back exactly what the phone was shown --
-            // including a partial one that was interrupted.
-            if let Err(err) = generate(&endpoint, &messages, &sampling, &cancel, &sink) {
-                let _ = sink.send(Event::Error {
-                    message: format!("{err:#}"),
-                });
-            }
-            let _ = sink.send(Event::Status {
-                state: SessionStatus::Idle,
-            });
-        });
-    }
-
-    fn answer_question(&self, _id: &str, _answers: &[String]) {
-        // Nothing here asks questions: this driver has no tools.
-    }
-
-    fn interrupt(&self) {
-        self.cancel.store(true, Ordering::Relaxed);
-    }
-
-    // Nothing to forward: this process has no notion of what the conversation
-    // is called, and the rename has already happened where the name lives.
-    fn set_title(&self, _title: &str) {}
-
-    fn set_permission_mode(&self, _mode: &str) {
-        let _ = self.sink.send(Event::Error {
-            message: "a llama.cpp session runs no tools, so there is nothing for a permission \
-                      mode to govern."
-                .to_string(),
-        });
-    }
-
-    fn set_model(&self, _model: &str) {
-        let _ = self.sink.send(Event::Error {
-            message: "a llama.cpp session's model is fixed when it starts, because the server \
-                      loads one model into memory. Spawn another session to use a different one."
-                .to_string(),
-        });
-    }
-
-    fn run_command(&self, text: &str) {
-        let _ = self.sink.send(Event::Error {
-            message: format!(
-                "a llama.cpp session has no commands of its own, so {text} means nothing to it."
-            ),
-        });
-    }
-
-    fn compact(&self) {
-        let _ = self.sink.send(Event::Error {
-            message: "llama.cpp has no compaction. Clear the session instead, which costs nothing."
-                .to_string(),
-        });
-    }
-
-    fn clear(&self) {
-        // All of it. `conversation` folds from the last of these, so recording
-        // the marker *is* the reset -- there is no driver state to keep in step
-        // with it, which is the same property that makes a second device see the
-        // same conversation this one does.
-        let _ = self.sink.send(Event::Cleared);
-    }
-
-    /// Stops generating and leaves the server loaded.
-    ///
-    /// Worth being deliberate about, because the cost points the other way from
-    /// the Claude driver's: a `llama-server` holds its whole model in memory, so
-    /// a leaked one is gigabytes nobody is using. It is left anyway, because the
-    /// alternative is unloading and reloading that model on every backend
-    /// restart -- minutes of disk, for a session somebody is in the middle of.
-    /// The record is what keeps it from being *nobody's*.
-    fn detach(&self) {
-        self.cancel.store(true, Ordering::Relaxed);
-    }
-
-    fn stop(&self) {
-        self.cancel.store(true, Ordering::Relaxed);
-        if let Some(record) = process::live(&self.session_dir) {
-            process::stop(&record, process::STOP_GRACE);
-        }
-        process::clear(&self.session_dir);
-    }
-}
-
-/// The conversation so far, folded out of the transcript.
-///
-/// Consecutive `AssistantText` deltas are one assistant turn, closed by the next
-/// user message -- which is also what makes an interrupted reply come back as
-/// the partial text the phone actually saw.
-///
-/// This must stay a pure function of the transcript and must never re-render
-/// earlier turns. llama.cpp caches the prompt prefix, so a growing conversation
-/// reprocesses almost nothing -- but only while every turn is byte-identical to
-/// last time. Changing how an old turn is rendered silently reprocesses the
-/// whole history on every message.
-fn conversation(path: &Path) -> Vec<Message> {
-    let Ok(events) = crate::session::transcript::read_after(path, 0) else {
-        return Vec::new();
-    };
-    let mut messages: Vec<Message> = Vec::new();
-    let mut pending = String::new();
-    // Everything before the last clear is still in the transcript and is
-    // deliberately not in the conversation. Folding from zero would put it back,
-    // which is the whole of what clearing had to undo.
-    let events = match events.iter().rposition(|e| e.event == Event::Cleared) {
-        Some(at) => &events[at + 1..],
-        None => &events[..],
-    };
-    for event in events.iter().cloned() {
-        match event.event {
-            Event::UserMessage { text, .. } => {
-                if !pending.is_empty() {
-                    messages.push(Message {
-                        role: "assistant".into(),
-                        content: std::mem::take(&mut pending),
-                    });
-                }
-                messages.push(Message {
-                    role: "user".into(),
-                    content: text,
-                });
-            }
-            Event::AssistantText { delta } => pending.push_str(&delta),
-            Event::AssistantTextFinal { text } => {
-                pending = text;
-            }
-            _ => {}
-        }
-    }
-    if !pending.is_empty() {
-        messages.push(Message {
-            role: "assistant".into(),
-            content: pending,
-        });
-    }
-    messages
-}
-
-/// Where a model key resolves to on disk, refusing anything that climbs
-/// out of the models directory -- the key arrives from a phone.
-fn model_path(models_dir: &Path, key: &str) -> Result<PathBuf> {
-    let mut path = models_dir.to_path_buf();
-    for part in key.split('/') {
-        if part.is_empty() || part == "." || part == ".." {
-            bail!("\"{key}\" is not a model key this can resolve");
-        }
-        path.push(part);
-    }
-    if !path.is_file() {
-        bail!("no downloaded model called \"{key}\" -- download it first");
-    }
-    Ok(path)
-}
-
-/// 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.
-///
-/// 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 {
-        if let Ok(response) = ureq::get(&url).call()
-            && response.status() == 200
-        {
-            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(),
-                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.
-fn generate(
-    endpoint: &str,
-    messages: &[Message],
-    sampling: &serde_json::Map<String, serde_json::Value>,
-    cancel: &AtomicBool,
-    sink: &EventSink,
-) -> Result<()> {
-    let mut body = json!({
-        "messages": messages,
-        "stream": true,
-        "stream_options": {"include_usage": true},
-    });
-    let map = body.as_object_mut().expect("built as an object");
-    for (key, value) in sampling {
-        map.insert(key.clone(), value.clone());
-    }
-
-    let mut response = ureq::post(format!("{endpoint}/v1/chat/completions"))
-        .header("Content-Type", "application/json")
-        .send_json(&body)
-        .context("asking llama-server to generate")?;
-
-    let reader = std::io::BufReader::new(response.body_mut().as_reader());
-    let mut tokens = 0u64;
-    // The prompt side only, which is what the model is holding -- the same
-    // definition the other dialects report, so one word on the phone means one
-    // thing whichever kind of session it is.
-    let mut context = None;
-    for line in std::io::BufRead::lines(reader) {
-        if cancel.load(Ordering::Relaxed) {
-            break;
-        }
-        let line = line.context("reading the generation stream")?;
-        // Server-sent events: the payload lines are the ones that matter.
-        let Some(payload) = line.strip_prefix("data: ") else {
-            continue;
-        };
-        if payload.trim() == "[DONE]" {
-            break;
-        }
-        let Ok(chunk) = serde_json::from_str::<serde_json::Value>(payload) else {
-            continue;
-        };
-        if let Some(usage) = chunk.get("usage") {
-            if let Some(total) = usage
-                .get("total_tokens")
-                .and_then(serde_json::Value::as_u64)
-            {
-                tokens = total;
-            }
-            if let Some(prompt) = usage
-                .get("prompt_tokens")
-                .and_then(serde_json::Value::as_u64)
-            {
-                context = Some(prompt);
-            }
-        }
-        let delta = chunk
-            .get("choices")
-            .and_then(|c| c.get(0))
-            .and_then(|c| c.get("delta"))
-            .and_then(|d| d.get("content"))
-            .and_then(serde_json::Value::as_str)
-            .unwrap_or_default();
-        if !delta.is_empty() {
-            let _ = sink.send(Event::AssistantText {
-                delta: delta.to_string(),
-            });
-        }
-    }
-    if tokens > 0 {
-        let _ = sink.send(Event::UsageDelta { tokens, context });
-    }
-    Ok(())
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use crate::session::transcript::Transcript;
-
-    /// Writes a transcript the way the pump does, so the fold is tested against
-    /// the real file format rather than a hand-built vector.
-    fn transcript_with(events: &[Event]) -> (tempfile::TempDir, PathBuf) {
-        let dir = tempfile::tempdir().expect("tempdir");
-        let path = dir.path().join("transcript.jsonl");
-        let mut transcript = Transcript::open(&path).expect("open");
-        for event in events {
-            transcript.append(event.clone(), 0.0).expect("append");
-        }
-        (dir, path)
-    }
-
-    #[test]
-    fn deltas_between_user_messages_are_one_assistant_turn() {
-        let (_dir, path) = transcript_with(&[
-            Event::UserMessage {
-                id: None,
-                text: "hello".into(),
-                attachments: Vec::new(),
-            },
-            Event::AssistantText {
-                delta: "hi ".into(),
-            },
-            Event::AssistantText {
-                delta: "there".into(),
-            },
-            Event::Status {
-                state: SessionStatus::Idle,
-            },
-            Event::UserMessage {
-                id: None,
-                text: "again".into(),
-                attachments: Vec::new(),
-            },
-            Event::AssistantText {
-                delta: "yes".into(),
-            },
-        ]);
-        let messages = conversation(&path);
-        assert_eq!(
-            messages
-                .iter()
-                .map(|m| (m.role.as_str(), m.content.as_str()))
-                .collect::<Vec<_>>(),
-            [
-                ("user", "hello"),
-                ("assistant", "hi there"),
-                ("user", "again"),
-                ("assistant", "yes")
-            ],
-        );
-    }
-
-    #[test]
-    /// The interrupted case, which decides what a resumed conversation is built
-    /// from: whatever the phone was shown. The deltas that arrived before the
-    /// stop are in the transcript, so they are in the prompt -- the model is
-    /// never told it said something the user did not see.
-    fn an_interrupted_reply_stays_in_the_conversation() {
-        let (_dir, path) = transcript_with(&[
-            Event::UserMessage {
-                id: None,
-                text: "count".into(),
-                attachments: Vec::new(),
-            },
-            Event::AssistantText {
-                delta: "one two".into(),
-            },
-            Event::Status {
-                state: SessionStatus::Idle,
-            },
-        ]);
-        let messages = conversation(&path);
-        assert_eq!(messages.len(), 2);
-        assert_eq!(messages[1].content, "one two");
-    }
-
-    #[test]
-    /// Events this driver does not produce must not disturb the fold: a
-    /// transcript can carry errors and status changes from a session that
-    /// was, say, relaunched.
-    fn other_events_are_not_part_of_the_conversation() {
-        let (_dir, path) = transcript_with(&[
-            Event::Status {
-                state: SessionStatus::Running,
-            },
-            Event::UserMessage {
-                id: None,
-                text: "hello".into(),
-                attachments: Vec::new(),
-            },
-            Event::Error {
-                message: "something went wrong".into(),
-            },
-            Event::AssistantText {
-                delta: "still here".into(),
-            },
-            Event::UsageDelta {
-                tokens: 12,
-                context: Some(12),
-            },
-        ]);
-        let messages = conversation(&path);
-        assert_eq!(messages.len(), 2);
-        assert_eq!(messages[0].content, "hello");
-        assert_eq!(messages[1].content, "still here");
-    }
-
-    #[test]
-    /// Clearing decides what the *model* is given, not just what the phone
-    /// draws. Everything above the marker stays in the transcript and none of it
-    /// is sent.
-    fn the_conversation_starts_after_the_last_clear() {
-        let (_dir, path) = transcript_with(&[
-            Event::UserMessage {
-                id: None,
-                text: "the long expensive conversation".into(),
-                attachments: Vec::new(),
-            },
-            Event::AssistantText {
-                delta: "at length".into(),
-            },
-            Event::Cleared,
-            Event::UserMessage {
-                id: None,
-                text: "a fresh start".into(),
-                attachments: Vec::new(),
-            },
-            Event::AssistantText {
-                delta: "cheaply".into(),
-            },
-        ]);
-        let messages = conversation(&path);
-        assert_eq!(messages.len(), 2);
-        assert_eq!(messages[0].content, "a fresh start");
-        assert_eq!(messages[1].content, "cheaply");
-    }
-
-    #[test]
-    /// The *last* one, so clearing twice does not resurrect what the
-    /// first clear dropped.
-    fn only_the_newest_clear_counts() {
-        let (_dir, path) = transcript_with(&[
-            Event::UserMessage {
-                id: None,
-                text: "one".into(),
-                attachments: Vec::new(),
-            },
-            Event::Cleared,
-            Event::UserMessage {
-                id: None,
-                text: "two".into(),
-                attachments: Vec::new(),
-            },
-            Event::Cleared,
-            Event::UserMessage {
-                id: None,
-                text: "three".into(),
-                attachments: Vec::new(),
-            },
-        ]);
-        let messages = conversation(&path);
-        assert_eq!(messages.len(), 1);
-        assert_eq!(messages[0].content, "three");
-    }
-
-    #[test]
-    fn a_model_key_cannot_climb_out_of_the_models_directory() {
-        let dir = tempfile::tempdir().expect("tempdir");
-        for attempt in ["../../etc/passwd", "unsloth/../../escape.gguf", ""] {
-            assert!(
-                model_path(dir.path(), attempt).is_err(),
-                "{attempt:?} should have been refused",
-            );
-        }
-    }
-}

server/src/session/llama/mcp.rs

@@ -0,0 +1,375 @@
+//! An MCP client, for the tools a llama session has that `llama-server` does
+//! not provide itself.
+//!
+//! **Why this is here and not a flag on `llama-server`.** That server can host
+//! MCP servers (`--mcp-servers-json`), but only ones it can *spawn*: its
+//! configuration is Cursor's, and an entry without a `command` is skipped with
+//! "MCP server 'exa' has no command". Exa's is a remote HTTP endpoint with
+//! nothing to spawn, so reaching it that way means a local process bridging
+//! stdio to HTTP -- a Node install on the machine serving the model, and a
+//! package to keep current, for what is three JSON-RPC calls.
+//!
+//! llama.cpp's own web UI does not do that either. It ships Exa in a
+//! "recommended servers" list and connects to `https://mcp.exa.ai/mcp`
+//! *itself*, from the browser. This is the same arrangement with this server
+//! in the browser's place, and it is the right one for a second reason: it
+//! puts the search on the machine running the backend rather than on whichever
+//! machine happens to be serving the model, which may have no route out at
+//! all.
+//!
+//! **Only the three calls a tool needs.** `initialize`, `tools/list`,
+//! `tools/call`. Nothing here implements resources, prompts, sampling or the
+//! server-to-client stream, because nothing here uses them; a session's tools
+//! are a list fetched once and a call made on demand. That is why this is a
+//! file rather than a dependency on a protocol crate -- there is no spec
+//! surface to get subtly wrong, only a request and its reply.
+//!
+//! Transport is "streamable HTTP": every message is a POST, and the reply is
+//! either JSON or a one-event SSE stream carrying the same JSON. Both are
+//! accepted because which one arrives is the server's choice, not ours.
+
+use anyhow::{Context, Result, bail};
+use serde_json::{Value, json};
+
+/// Identifies this client to an MCP server.
+///
+/// Not politeness: Exa's endpoint is behind Cloudflare, which answers **403**
+/// to a request with no `User-Agent` at all (measured 2026-09-19 -- the same
+/// request with one succeeds). A client that omitted it would look exactly
+/// like a server that was refusing us.
+const USER_AGENT: &str = concat!("ai-server/", env!("CARGO_PKG_VERSION"));
+
+/// The protocol version this speaks. Sent at `initialize`; a server that
+/// prefers another says so in its answer and this goes along with whatever it
+/// then sends, since none of the three calls here has changed between
+/// versions.
+const PROTOCOL_VERSION: &str = "2025-06-18";
+
+/// How long any one call may take.
+///
+/// Generous because a web search is a search: Exa fetches and cleans pages
+/// before answering. Bounded at all because this blocks a turn, and a tool
+/// that never returns is a session that never speaks again.
+const CALL_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(120);
+
+/// A connected MCP server, and the tools it offered.
+pub struct McpServer {
+    /// The name this server is configured under. It prefixes every tool, so
+    /// two servers offering `search` are two different tools.
+    name: String,
+    url: String,
+    /// What the server called this conversation, when it named one. Sent back
+    /// on every later request; a server that keeps no session sends no header
+    /// and this stays `None`.
+    session: Option<String>,
+    /// The tool names this server answers to, without the prefix, keyed by the
+    /// prefixed name the model is given.
+    tools: Vec<McpTool>,
+}
+
+/// One tool an MCP server offers, in both the names it has.
+pub struct McpTool {
+    /// `{server}_{tool}` -- what the model calls it, and what comes back in a
+    /// tool call. Prefixed the way `llama-server` prefixes the MCP tools it
+    /// hosts itself, so a reader sees one naming convention whichever side a
+    /// tool came from.
+    pub qualified: String,
+    /// What the server calls it.
+    bare: String,
+    /// The OpenAI-shaped function definition sent to the model.
+    pub definition: Value,
+}
+
+impl McpServer {
+    /// Connects, handshakes, and asks what it can do.
+    ///
+    /// All three steps or none: a server that answered `initialize` and then
+    /// failed to list its tools is not a server with no tools, and returning
+    /// an empty list for it would put a session on screen that silently
+    /// cannot search.
+    pub fn connect(name: &str, url: &str) -> Result<Self> {
+        let mut server = Self {
+            name: name.to_string(),
+            url: url.to_string(),
+            session: None,
+            tools: Vec::new(),
+        };
+        server
+            .request(
+                1,
+                "initialize",
+                json!({
+                    "protocolVersion": PROTOCOL_VERSION,
+                    "capabilities": {},
+                    "clientInfo": {"name": "ai-server", "title": "AI Sessions", "version": env!("CARGO_PKG_VERSION")},
+                }),
+            )
+            .with_context(|| format!("handshaking with the {name} MCP server at {url}"))?;
+        // A notification: no id, and the server answers with no body. Sent
+        // because the specification requires it before any other call, and
+        // Exa's server does enforce it.
+        server.notify("notifications/initialized")?;
+        let listed = server
+            .request(2, "tools/list", json!({}))
+            .with_context(|| format!("asking the {name} MCP server what it offers"))?;
+        server.tools = listed
+            .get("tools")
+            .and_then(Value::as_array)
+            .map(|tools| {
+                tools
+                    .iter()
+                    .filter_map(|tool| server.describe(tool))
+                    .collect()
+            })
+            .unwrap_or_default();
+        Ok(server)
+    }
+
+    /// Turns one entry of `tools/list` into the function definition a model is
+    /// given, or `None` for one this cannot name or call.
+    fn describe(&self, tool: &Value) -> Option<McpTool> {
+        let bare = tool.get("name").and_then(Value::as_str)?.to_string();
+        let qualified = format!("{}_{bare}", self.name);
+        let mut function = serde_json::Map::new();
+        function.insert("name".into(), json!(qualified));
+        if let Some(description) = tool.get("description").and_then(Value::as_str) {
+            function.insert("description".into(), json!(description));
+        }
+        // `inputSchema` in MCP, `parameters` in the OpenAI shape: the same
+        // JSON Schema under two names. A tool that declares none takes no
+        // arguments, which is an empty object rather than an absent key --
+        // some templates render the key unconditionally.
+        function.insert(
+            "parameters".into(),
+            tool.get("inputSchema")
+                .cloned()
+                .unwrap_or_else(|| json!({"type": "object", "properties": {}})),
+        );
+        Some(McpTool {
+            qualified,
+            bare,
+            definition: json!({"type": "function", "function": function}),
+        })
+    }
+
+    pub fn tools(&self) -> &[McpTool] {
+        &self.tools
+    }
+
+    /// Runs one of this server's tools, named as the model named it.
+    ///
+    /// The result is the text a model is shown. A tool the server reports as
+    /// failing is **not** an error here: `isError` means the tool ran and went
+    /// wrong -- a search that found nothing, a page that would not fetch --
+    /// and the model is the one that has to know, so it comes back as its own
+    /// message. An error is reserved for not having reached the server at all.
+    pub fn call(&mut self, qualified: &str, arguments: &Value) -> Result<String> {
+        // The bare name is taken before the call, because the call needs the
+        // whole of `self` and the tool list is part of it.
+        let bare = self
+            .tools
+            .iter()
+            .find(|tool| tool.qualified == qualified)
+            .map(|tool| tool.bare.clone())
+            .with_context(|| format!("{} does not offer {qualified}", self.name))?;
+        let result = self.request(
+            3,
+            "tools/call",
+            json!({"name": bare, "arguments": arguments}),
+        )?;
+        Ok(rendered(&result))
+    }
+
+    /// One request, and its result.
+    ///
+    /// `&self` rather than `&mut self` everywhere but the handshake would be
+    /// tidier and is wrong: the session header is assigned by the server on
+    /// the first reply and has to be kept.
+    fn request(&mut self, id: u64, method: &str, params: Value) -> Result<Value> {
+        let body = json!({"jsonrpc": "2.0", "id": id, "method": method, "params": params});
+        let answer = self.post(&body)?.with_context(|| {
+            format!(
+                "the {} MCP server answered {method} with nothing",
+                self.name
+            )
+        })?;
+        if let Some(message) = answer.pointer("/error/message").and_then(Value::as_str) {
+            bail!("{} refused {method}: {message}", self.name);
+        }
+        answer
+            .get("result")
+            .cloned()
+            .with_context(|| format!("the {} MCP server's {method} carried no result", self.name))
+    }
+
+    /// A message with no id, which is answered with no body.
+    fn notify(&mut self, method: &str) -> Result<()> {
+        self.post(&json!({"jsonrpc": "2.0", "method": method}))?;
+        Ok(())
+    }
+
+    /// Posts one JSON-RPC message and returns whatever came back, which for a
+    /// notification is nothing.
+    fn post(&mut self, body: &Value) -> Result<Option<Value>> {
+        let mut request = ureq::post(&self.url)
+            .config()
+            .timeout_global(Some(CALL_TIMEOUT))
+            .build()
+            .header("content-type", "application/json")
+            // Both, because which one a server replies with is its choice.
+            .header("accept", "application/json, text/event-stream")
+            .header("user-agent", USER_AGENT);
+        if let Some(session) = &self.session {
+            request = request.header("mcp-session-id", session);
+        }
+        let mut response = request
+            .send_json(body)
+            .with_context(|| format!("reaching the {} MCP server at {}", self.name, self.url))?;
+        if let Some(session) = response
+            .headers()
+            .get("mcp-session-id")
+            .and_then(|value| value.to_str().ok())
+        {
+            self.session = Some(session.to_string());
+        }
+        let streamed = response
+            .headers()
+            .get("content-type")
+            .and_then(|value| value.to_str().ok())
+            .is_some_and(|value| value.contains("text/event-stream"));
+        let text = response
+            .body_mut()
+            .read_to_string()
+            .with_context(|| format!("reading the {} MCP server's answer", self.name))?;
+        Ok(first_message(&text, streamed))
+    }
+}
+
+/// The first JSON-RPC message in a reply body.
+///
+/// One, not all: every call here carries a single id and the server answers it
+/// once. Server-sent events are unwrapped to their payload lines; a plain JSON
+/// body is itself.
+fn first_message(text: &str, streamed: bool) -> Option<Value> {
+    if streamed {
+        return text
+            .lines()
+            .filter_map(|line| line.strip_prefix("data: "))
+            .find_map(|payload| serde_json::from_str(payload).ok());
+    }
+    serde_json::from_str(text.trim()).ok()
+}
+
+/// A `tools/call` result as the text a model is given.
+///
+/// MCP answers with a list of content blocks; the text ones are joined and the
+/// rest are named rather than dropped, because a model told nothing came back
+/// will try again. `structuredContent` is used when there is no text at all,
+/// which is how some servers answer entirely.
+fn rendered(result: &Value) -> String {
+    let blocks = result.get("content").and_then(Value::as_array);
+    let mut parts: Vec<String> = Vec::new();
+    for block in blocks.into_iter().flatten() {
+        match block.get("type").and_then(Value::as_str) {
+            Some("text") => parts.push(
+                block
+                    .get("text")
+                    .and_then(Value::as_str)
+                    .unwrap_or_default()
+                    .to_string(),
+            ),
+            Some(kind) => parts.push(format!("[{kind} content, which this session cannot show]")),
+            None => {}
+        }
+    }
+    if parts.iter().all(|part| part.trim().is_empty())
+        && let Some(structured) = result.get("structuredContent")
+    {
+        return structured.to_string();
+    }
+    parts.join("\n")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn an_event_stream_body_is_unwrapped_to_its_payload() {
+        let body =
+            "event: message\ndata: {\"jsonrpc\":\"2.0\",\"id\":1,\"result\":{\"ok\":true}}\n\n";
+        assert_eq!(
+            first_message(body, true),
+            Some(json!({"jsonrpc": "2.0", "id": 1, "result": {"ok": true}})),
+        );
+    }
+
+    #[test]
+    fn a_plain_json_body_is_the_message() {
+        let body = "  {\"jsonrpc\":\"2.0\",\"id\":1,\"result\":{}}\n";
+        assert_eq!(
+            first_message(body, false),
+            Some(json!({"jsonrpc": "2.0", "id": 1, "result": {}})),
+        );
+    }
+
+    #[test]
+    /// A notification's reply, which is nothing at all.
+    fn an_empty_body_is_no_message() {
+        assert_eq!(first_message("", false), None);
+        assert_eq!(first_message("event: ping\n", true), None);
+    }
+
+    #[test]
+    fn text_blocks_are_joined_and_other_kinds_are_named() {
+        let result = json!({"content": [
+            {"type": "text", "text": "first"},
+            {"type": "image", "data": "…"},
+            {"type": "text", "text": "second"},
+        ]});
+        assert_eq!(
+            rendered(&result),
+            "first\n[image content, which this session cannot show]\nsecond",
+        );
+    }
+
+    #[test]
+    /// A server that answers only in structured form. Rendering "" for it
+    /// would tell the model the search came back empty, which is a different
+    /// fact from the one that is true.
+    fn a_result_with_no_text_falls_back_to_its_structured_form() {
+        let result = json!({"content": [], "structuredContent": {"hits": 2}});
+        assert_eq!(rendered(&result), "{\"hits\":2}");
+    }
+
+    #[test]
+    /// The real endpoint, which is the only thing that can confirm the
+    /// handshake, the session header and the SSE unwrapping all agree with a
+    /// server nobody here wrote. Skipped without network rather than failed:
+    /// `./run-tests.sh` has to pass on a machine with no route out.
+    fn exa_answers_a_search_over_the_real_protocol() {
+        let Ok(mut server) = McpServer::connect("exa", super::super::EXA_MCP_URL) else {
+            eprintln!("skipping: could not reach Exa");
+            return;
+        };
+        assert!(
+            server
+                .tools()
+                .iter()
+                .any(|tool| tool.qualified == "exa_web_search_exa"),
+            "Exa offered {:?}",
+            server
+                .tools()
+                .iter()
+                .map(|tool| &tool.qualified)
+                .collect::<Vec<_>>(),
+        );
+        let answer = server
+            .call(
+                "exa_web_search_exa",
+                &json!({"query": "llama.cpp server", "numResults": 1}),
+            )
+            .expect("search");
+        assert!(!answer.trim().is_empty(), "a search returned nothing");
+    }
+}

server/src/session/llama/tools.rs

@@ -0,0 +1,263 @@
+//! What a llama session can do besides talk, and who runs it.
+//!
+//! Two sources, one list. `llama-server` started with `--tools` runs a set of
+//! its own -- reading, searching, editing, a shell -- and publishes them at
+//! `GET /tools` in the shape a model is given, with `POST /tools` to run one.
+//! Anything else comes from an MCP server this backend is connected to (see
+//! [`super::mcp`]). Both arrive here as a definition to offer and a way to
+//! call, and nothing downstream of [`Tools::execute`] knows which a tool was.
+//!
+//! **The built-in tools run where the model does, and that is the point.** A
+//! session on another machine edits that machine's files, because that is the
+//! machine `llama-server` is on -- the same rule the model file already
+//! follows. MCP tools run here instead, which is right for the opposite
+//! reason: a web search wants the machine with a route out, not the one with
+//! the GPU.
+//!
+//! **A tool's failure is a result, not an error.** A missing file, a command
+//! that exited non-zero, a search that found nothing: all of those are things
+//! the model has to read and act on, so they come back as the tool's output.
+//! [`Tools::execute`] returns `Err` only when the tool could not be reached at
+//! all, which is a fact about this server rather than about the work.
+
+use std::collections::HashMap;
+use std::sync::{Arc, Mutex};
+
+use anyhow::{Context, Result};
+use serde_json::{Value, json};
+
+use super::mcp::McpServer;
+
+/// How long one tool call may take.
+///
+/// This is the shell tool's budget as much as anything: a build, a test run,
+/// a `find` over a large tree. Bounded because it blocks the turn, and a
+/// session stuck behind a command that will never finish cannot even be told
+/// to stop.
+const EXECUTE_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(300);
+
+/// The tools one session has, and how to run each of them.
+pub struct Tools {
+    /// The `llama-server` these belong to. Replaced when the session's model
+    /// changes, because that is a different server on a different port.
+    endpoint: String,
+    /// What the model is given, in the order it is offered: the server's own
+    /// tools first, then each MCP server's.
+    definitions: Vec<Value>,
+    /// The server's tools, and whether each is run relative to a working
+    /// directory. Only the ones that say so are sent one -- a tool that
+    /// ignores it would still have its cache keyed on it.
+    server: HashMap<String, bool>,
+    /// Connected MCP servers, each of which knows its own tools by the
+    /// prefixed names they were offered under.
+    mcp: Vec<Arc<Mutex<McpServer>>>,
+}
+
+impl Tools {
+    /// Asks a ready `llama-server` what it offers and adds what the MCP
+    /// servers offered.
+    ///
+    /// A server started without `--tools` answers with an empty list, and a
+    /// session with only MCP tools is a perfectly good session -- so nothing
+    /// here treats "no tools" as a failure. What *is* a failure is not being
+    /// able to ask, because that is the same server the conversation is about
+    /// to go to.
+    pub fn discover(endpoint: &str, mcp: Vec<Arc<Mutex<McpServer>>>) -> Result<Self> {
+        let catalog: Vec<Value> = ureq::get(format!("{endpoint}/tools"))
+            .call()
+            .context("asking llama-server which tools it has")?
+            .body_mut()
+            .read_json()
+            .context("reading llama-server's tool list")?;
+        let mut definitions = Vec::new();
+        let mut server = HashMap::new();
+        for entry in &catalog {
+            let Some(name) = entry
+                .pointer("/definition/function/name")
+                .and_then(Value::as_str)
+            else {
+                continue;
+            };
+            let Some(definition) = entry.get("definition") else {
+                continue;
+            };
+            server.insert(
+                name.to_string(),
+                entry
+                    .get("uses_cwd")
+                    .and_then(Value::as_bool)
+                    .unwrap_or(false),
+            );
+            definitions.push(definition.clone());
+        }
+        for connected in &mcp {
+            for tool in connected.lock().unwrap().tools() {
+                definitions.push(tool.definition.clone());
+            }
+        }
+        Ok(Self {
+            endpoint: endpoint.to_string(),
+            definitions,
+            server,
+            mcp,
+        })
+    }
+
+    /// What goes in the request's `tools`, or `None` when there is nothing to
+    /// offer.
+    ///
+    /// Absent rather than empty for a reason that shows on screen: a chat
+    /// template branches on whether tools were given, and an empty list
+    /// renders the whole "you may call one or more functions" preamble with no
+    /// functions under it.
+    pub fn offered(&self) -> Option<&[Value]> {
+        (!self.definitions.is_empty()).then_some(&self.definitions)
+    }
+
+    /// Whether this is a tool at all, which decides what to do about a call
+    /// naming something else.
+    pub fn knows(&self, name: &str) -> bool {
+        self.server.contains_key(name) || self.mcp_for(name).is_some()
+    }
+
+    /// The MCP server that offered `name`, if one did.
+    fn mcp_for(&self, name: &str) -> Option<&Arc<Mutex<McpServer>>> {
+        self.mcp.iter().find(|server| {
+            server
+                .lock()
+                .unwrap()
+                .tools()
+                .iter()
+                .any(|tool| tool.qualified == name)
+        })
+    }
+
+    /// Runs one call and returns what the model should read.
+    ///
+    /// `cwd` is the session's working directory, sent only to the tools that
+    /// say they use one. A session with no working directory sends none, and
+    /// `llama-server` falls back to its own -- which is the honest outcome:
+    /// this server has no better answer for where "here" is.
+    pub fn execute(&self, name: &str, arguments: &Value, cwd: Option<&str>) -> Result<String> {
+        if let Some(server) = self.mcp_for(name) {
+            return server.lock().unwrap().call(name, arguments);
+        }
+        let uses_cwd = *self
+            .server
+            .get(name)
+            .with_context(|| format!("no tool called {name}"))?;
+        let mut request = ureq::post(format!("{}/tools", self.endpoint))
+            .config()
+            .timeout_global(Some(EXECUTE_TIMEOUT))
+            // The refusal is a sentence the model can act on, so it is read as
+            // one rather than discarded in favour of its status code -- the
+            // same reason `super::refusal` exists for generation.
+            .http_status_as_error(false)
+            .build()
+            .header("content-type", "application/json");
+        if let (true, Some(cwd)) = (uses_cwd, cwd) {
+            request = request.header("x-tool-cwd", cwd);
+        }
+        let mut response = request
+            .send_json(json!({"tool": name, "params": arguments}))
+            .with_context(|| format!("asking llama-server to run {name}"))?;
+        let body = response
+            .body_mut()
+            .read_to_string()
+            .with_context(|| format!("reading what {name} produced"))?;
+        Ok(match serde_json::from_str::<Value>(&body) {
+            Ok(answer) => result_text(&answer),
+            // Not JSON at all: hand over what was said rather than a parse
+            // error about it, since the model is what has to carry on.
+            Err(_) => body,
+        })
+    }
+}
+
+/// `POST /tools`'s answer as the text a model is given.
+///
+/// The server answers `plain_text_response` for a tool that ran and `error`
+/// for one that did not, and both are the model's business -- see this
+/// module's note on failures being results. Anything else is handed over as
+/// itself rather than discarded, since a tool this build has not seen before
+/// is exactly the case where guessing is worst.
+fn result_text(answer: &Value) -> String {
+    if let Some(text) = answer.get("plain_text_response").and_then(Value::as_str) {
+        return text.to_string();
+    }
+    if let Some(message) = answer.get("error").and_then(Value::as_str) {
+        return message.to_string();
+    }
+    answer.to_string()
+}
+
+/// How much a session asks before it acts.
+///
+/// Two, because two is what the mechanism underneath actually has. The web UI
+/// that ships with `llama-server` asks before every call and remembers the
+/// tools you said "always" to, and that pair -- a prompt and a growing set of
+/// exceptions -- is the whole of its permission model. A third mode sitting
+/// between them would have to invent a rule about which tools are "edits",
+/// and the rule would be this app's opinion rather than anything the tools
+/// declare.
+pub const MODES: &[&str] = &["manual", "bypassPermissions"];
+
+/// What a new llama session asks by default.
+///
+/// The cautious one, matching the web UI: a model with a shell on somebody's
+/// own machine is the case to be wrong about in this direction, and one tap
+/// on "always allow" is what makes it bearable afterwards.
+pub const DEFAULT_MODE: &str = "manual";
+
+/// The answer that makes an allowance permanent for the session. The tool's
+/// name follows it, which is what makes the transcript alone enough to
+/// rebuild the set -- see `super::allowed`.
+pub const ALWAYS_PREFIX: &str = "Always allow ";
+pub const ALLOW_ONCE: &str = "Allow once";
+pub const REFUSE: &str = "Don't allow";
+
+/// What the model is told when a call was refused.
+///
+/// Addressed to the model, not to the reader: it has to understand that the
+/// work did not happen and that trying the same call again is not the way
+/// round it, or it retries in a loop.
+pub const REFUSED: &str = "The person using this session did not allow this call, so it was not run. Do not try it \
+     again -- say what you were going to do and why it needed that, and let them decide.";
+
+/// What stands in for a call that never finished, when the transcript is read
+/// back into a conversation.
+///
+/// Every tool call in the history owes a result, because that is the shape a
+/// chat template renders; a turn stopped between the call and its result
+/// leaves one that has none. Saying so is better than inventing an outcome,
+/// and better than dropping the call -- which would tell the model it never
+/// asked.
+pub const UNFINISHED: &str = "This call was interrupted before it produced anything.";
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn a_tool_that_ran_reads_as_its_output() {
+        assert_eq!(
+            result_text(&json!({"plain_text_response": "hello\n"})),
+            "hello\n",
+        );
+    }
+
+    #[test]
+    /// The model is told what went wrong, because the model is what has to do
+    /// something about it -- read a different path, fix the command.
+    fn a_tool_that_failed_reads_as_its_message() {
+        assert_eq!(
+            result_text(&json!({"error": "cannot stat file: /tmp/nope"})),
+            "cannot stat file: /tmp/nope",
+        );
+    }
+
+    #[test]
+    fn anything_else_is_handed_over_as_itself() {
+        assert_eq!(result_text(&json!({"rows": 2})), "{\"rows\":2}");
+    }
+}

server/src/session/mod.rs

@@ -3980,6 +3980,7 @@ mod tests {
                     kind: DriverKind::ClaudeCli,
                     command: Some(command.to_string_lossy().into_owned()),
                     models: Vec::new(),
+                    mcp_servers: Vec::new(),
                 },
             ])],
             ..Config::default()

server/src/usage.rs

@@ -1108,6 +1108,7 @@ mod tests {
                 kind: DriverKind::ClaudeCli,
                 command: None,
                 models: vec![],
+                mcp_servers: Vec::new(),
             }],
         }
     }
@@ -1171,6 +1172,7 @@ mod tests {
             kind: DriverKind::Echo,
             command: None,
             models: vec![],
+            mcp_servers: Vec::new(),
         }];
         // A machine with no Claude on it has no Claude limits, and a row
         // reporting on it would be a fact about nothing. Echo included: