Declare provider settings, and give the context figure a denominator

Two things a session could not say, and one it was saying wrongly.

**Every provider setting is reachable.** `-np 1`, the MTP draft depth, the
tool set, the sampling parameters -- most were hardcoded to what measured
best on this machine, which is right as a default and wrong as a constant:
the next machine has a different GPU and a different core count, and
nobody running this app can edit the source. `DriverKind::params` now
declares what a provider takes -- key, label, shape, what blank means, and
whether a change waits for a restart -- and the phone renders whatever
arrives, on the spawn form and in the session settings dialog. Adding a
setting to a driver is one entry in that table and no app change.
`POST /sessions/{id}/params` takes the whole map, so an absent key is the
instruction to unset; the sampling half applies at once and the session is
told in words which of the rest are waiting for a restart.

`tools` is one of them, because it is the biggest lever on a tight
context: the seven built-in definitions are ~1,300 tokens of every prompt
(2,191 against 887 with none). `"none"` omits the flag rather than passing
it on, since `--tools none` is `unknown tool "none"` and a server that
exits.

**The context figure has a denominator.** `Event::ContextWindow` carries
it, read from `llama-server`'s `/props` once the model is up -- the
measurement rather than the request, since a session that named no context
size gets the model's own. Neither coding CLI states its window, so those
keep the bare figure: "2,042" and "2,042 / 8,192" are deliberately
different-looking, and a missing ceiling is never drawn as a proportion of
an assumed one.

**And the numerator was wrong**, by the length of the last reply: it was
the prompt alone, so a five-word answer reported 2,042 against a slot
holding 2,355. It is the turn's total now, which matches `llama-server`'s
own `n_tokens` to within a token.

Two defects the review found, both of which would have shipped: changing
settings on a *stopped* session reported "no process running, so it can't
take new settings", when a stopped session is exactly when you would set
them for the next start; and `GET /tools` answers **403** rather than an
empty list on a server started without `--tools`, so reading it as a
failure made the no-tools session one that never started.

Verified against real models: settings spawned and changed live, the
restart note, a session with two tools and one with none, and the counter
checked against the server's own slot occupancy each time.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

AGENTS.md

@@ -62,6 +62,15 @@ Module-by-module intent is in PLAN.md's "Backend layout".
   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**.
+  **Every one of those is a default rather than a constant** (2026-09-19):
+  `DriverKind::params` declares what a provider takes — key, label, shape,
+  and whether a change waits for a restart — and the phone renders whatever
+  arrives, on the spawn form and in the session settings dialog. Adding a
+  setting to a driver is one entry in that table and no app change. `tools`
+  is in there too, because the seven built-in definitions are ~1,300 tokens
+  of every prompt (2,191 against 887 with none), which on a small window is
+  the difference between a usable session and one that overruns; `"none"`
+  omits the flag, since `--tools none` is a server that exits.
   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
@@ -325,6 +334,11 @@ written, and the fold uses that same predicate to decide a reply is settled.
 
 ## Things that have bitten
 
+- **A server started with no `--tools` answers 403 at `GET /tools`, not an
+  empty list.** The route is off rather than empty, so reading that as a
+  failure made "no tools" — the one setting whose entire purpose is to have
+  none — a session that never started.
+
 - **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

PLAN.md

@@ -1435,6 +1435,26 @@ 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.
+- **What a provider's settings are is declared by the server** (2026-09-19,
+  `DriverKind::params`). A spec is a key, words, a shape and whether a change
+  waits for a restart; the phone renders whatever arrives, so a driver that
+  grows a setting gets a control with no app change. The reason it is
+  declared rather than drawn is not tidiness: several `llama-server` flags
+  were hardcoded to what measured best on one machine, which is right as a
+  default and wrong as a constant — the next machine has a different GPU, and
+  nobody running this app can edit the source. `POST /sessions/{id}/params`
+  takes the whole map, so an absent key *is* the instruction to unset.
+- **The context figure has a denominator where one can be measured**
+  (2026-09-19). `Event::ContextWindow` carries it, read from `llama-server`'s
+  `/props` once the model is up — the measurement rather than the request,
+  since a session that named no context size gets the model's own. Neither
+  coding CLI states its window, so those keep the bare figure: "2,042" and
+  "2,042 / 8,192" are deliberately different-looking, and a missing ceiling is
+  never drawn as a proportion of an assumed one.
+  The numerator was also wrong, by the length of the last reply: it was the
+  prompt alone, so a five-word answer reported 2,042 against a slot holding
+  2,355. It is the turn's total now, which matches `llama-server`'s own
+  `n_tokens` exactly.
 - **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

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

@@ -204,6 +204,13 @@ data class SessionSummary(
      * full.
      */
     val contextTokens: Long?,
+    /**
+     * What [contextTokens] is out of, or null where this session's provider does not say. A third
+     * state, not a fourth reading of the same one: the occupancy is known and the ceiling is not.
+     */
+    val contextLimit: Long?,
+    /** What this session's provider settings are set to; empty where it takes none. */
+    val params: Map<String, String>,
     /**
      * The longest edge an image should have when it reaches this session, or null where the
      * provider has no limit.
@@ -262,6 +269,8 @@ private fun parseSession(session: JSONObject) =
         cwd = session.optString("cwd").ifEmpty { null },
         contextTokens =
             if (session.has("contextTokens")) session.getLong("contextTokens") else null,
+        contextLimit = if (session.has("contextLimit")) session.getLong("contextLimit") else null,
+        params = session.optJSONObject("params").stringMap(),
         maxImageEdge = session.optInt("maxImageEdge", 0).takeIf { it > 0 },
         usageProvider = session.optString("usageProvider").ifEmpty { null },
         status = session.getString("status"),
@@ -344,6 +353,27 @@ data class Provider(
     val models: List<String>,
     val permissionModes: List<String>,
     val defaultPermissionMode: String?,
+    /** The extra settings this provider takes, in the order to draw them. See [ParamSpec]. */
+    val params: List<ParamSpec>,
+)
+
+/**
+ * One setting a provider takes, as the server describes it.
+ *
+ * Declared by the server rather than drawn here, so a driver that grows a setting gets a control
+ * with no app change — and, more to the point, so that a value measured on one machine can ship as
+ * a default without becoming a constant nobody else can reach.
+ */
+data class ParamSpec(
+    val key: String,
+    val label: String,
+    /** What leaving it blank means, in words, shown as the placeholder. */
+    val unset: String,
+    val kind: String,
+    /** For [kind] `"choice"`: the options, the first of which means "unset". */
+    val options: List<String>,
+    /** Whether a change waits for the session's process to start again. */
+    val restart: Boolean,
 )
 
 /**
@@ -359,6 +389,11 @@ data class Machine(
     val providers: List<Provider>,
 )
 
+/** A JSON object of strings, and the empty map for one that is absent. */
+private fun JSONObject?.stringMap(): Map<String, String> =
+    this?.let { object_ -> object_.keys().asSequence().associateWith { object_.getString(it) } }
+        ?: emptyMap()
+
 private fun parseProvider(provider: JSONObject): Provider {
     val kind = provider.getString("kind")
     return Provider(
@@ -368,6 +403,17 @@ private fun parseProvider(provider: JSONObject): Provider {
         models = provider.optJSONArray("models")?.strings().orEmpty(),
         permissionModes = provider.optJSONArray("permissionModes")?.strings().orEmpty(),
         defaultPermissionMode = provider.optString("defaultPermissionMode").ifEmpty { null },
+        params =
+            provider.optJSONArray("params")?.mapObjects { spec ->
+                ParamSpec(
+                    key = spec.getString("key"),
+                    label = spec.getString("label"),
+                    unset = spec.getString("unset"),
+                    kind = spec.getString("kind"),
+                    options = spec.optJSONArray("options")?.strings().orEmpty(),
+                    restart = spec.optBoolean("restart"),
+                )
+            } ?: emptyList(),
     )
 }
 
@@ -481,6 +527,13 @@ data class Importable(
      * history from before a compaction, which the model is no longer given.
      */
     val contextTokens: Long?,
+    /**
+     * What [contextTokens] is out of, or null where this session's provider does not say. A third
+     * state, not a fourth reading of the same one: the occupancy is known and the ceiling is not.
+     */
+    val contextLimit: Long?,
+    /** What this session's provider settings are set to; empty where it takes none. */
+    val params: Map<String, String>,
     /** Whether [title] is a name somebody chose rather than the last thing said in the session. */
     val named: Boolean,
     /**
@@ -555,6 +608,10 @@ fun fetchImportable(settings: ServerSettings, machine: String): List<Importable>
                 contextTokens =
                     if (session.isNull("contextTokens")) null
                     else session.optLong("contextTokens").takeIf { it > 0L },
+                contextLimit =
+                    if (session.isNull("contextLimit")) null
+                    else session.optLong("contextLimit").takeIf { it > 0L },
+                params = session.optJSONObject("params").stringMap(),
                 // Absent means an older backend that cannot answer, which is what "unknown" says.
                 inUse = session.optString("inUse", "unknown"),
                 named = session.optBoolean("named", false),
@@ -1240,6 +1297,16 @@ fun setSessionEffort(settings: ServerSettings, sessionId: String, level: String?
 }
 
 /** Switches how much a running session asks before acting, also in place. */
+/** Replaces a session's provider settings with [params] — the whole map, not a patch. */
+fun setSessionParams(settings: ServerSettings, sessionId: String, params: Map<String, String>) {
+    requestFromServer(
+        settings,
+        "/sessions/$sessionId/params",
+        method = "POST",
+        jsonBody = JSONObject().put("params", JSONObject(params.toMap())).toString(),
+    ) {}
+}
+
 fun setSessionPermissionMode(settings: ServerSettings, sessionId: String, mode: String) {
     requestFromServer(
         settings,

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

@@ -60,3 +60,23 @@ fun compactingLabel(seconds: Long?): String =
         seconds < 60 -> "compacting ${seconds}s"
         else -> "compacting ${seconds / 60}m ${seconds % 60}s"
     }
+
+/**
+ * How full the session is, as the status row says it.
+ *
+ * Three states, not two, and the third is the one that needed the words: a session whose occupancy
+ * is known and whose ceiling is not. That one keeps the bare figure, and a session with a ceiling
+ * gets both — the reader can see which they are looking at. What must not happen is a missing
+ * ceiling drawn as a number, or as a proportion of some assumed window, which would be this screen
+ * inventing the very fact it does not have.
+ *
+ * A llama.cpp session always has one, since the window is a flag its own server was started with. A
+ * coding CLI's is the vendor's business and neither control protocol states it, so those keep the
+ * bare figure they have always had.
+ */
+fun contextLabel(held: Long?, limit: Long?): String =
+    when {
+        held == null -> "context unknown"
+        limit == null -> "context ${tokens(held)}"
+        else -> "context ${tokens(held)} / ${tokens(limit)}"
+    }

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

@@ -163,6 +163,9 @@ sealed class SessionEvent {
      */
     data class UsageDelta(val tokens: Long, val context: Long?) : SessionEvent()
 
+    /** How much context this session's model has, which is what [UsageDelta.context] is out of. */
+    data class ContextWindow(val tokens: Long) : SessionEvent()
+
     /**
      * A compaction that finished, and how much context it recovered.
      *
@@ -293,6 +296,7 @@ fun parseSeqEvent(json: String): SeqEvent {
                     model = body.optString("model").ifEmpty { null },
                     permissionMode = body.optString("permissionMode").ifEmpty { null },
                 )
+            "contextWindow" -> SessionEvent.ContextWindow(body.getLong("tokens"))
             "usageDelta" ->
                 SessionEvent.UsageDelta(
                     body.getLong("tokens"),
@@ -363,3 +367,18 @@ fun contextAfter(current: Long?, event: SessionEvent): Long? =
         is SessionEvent.Cleared -> null
         else -> current
     }
+
+/**
+ * The context window after [event], mirroring the server's `context_limit_after` for the same
+ * reason [contextAfter] mirrors its neighbour: the screen has to keep up between page loads.
+ *
+ * A window belongs to the process, so a session whose process has exited has none — left standing,
+ * a session restarted on a different model would draw its occupancy against the old model's
+ * ceiling.
+ */
+fun contextLimitAfter(current: Long?, event: SessionEvent): Long? =
+    when (event) {
+        is SessionEvent.ContextWindow -> event.tokens
+        is SessionEvent.Status -> if (event.state == "exited") null else current
+        else -> current
+    }

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

@@ -0,0 +1,117 @@
+package com.example.aiapp
+
+import androidx.compose.foundation.layout.Column
+import androidx.compose.foundation.layout.Spacer
+import androidx.compose.foundation.layout.fillMaxWidth
+import androidx.compose.foundation.layout.height
+import androidx.compose.foundation.text.KeyboardOptions
+import androidx.compose.material3.MaterialTheme
+import androidx.compose.material3.OutlinedTextField
+import androidx.compose.material3.Text
+import androidx.compose.runtime.Composable
+import androidx.compose.ui.Modifier
+import androidx.compose.ui.text.input.KeyboardType
+import androidx.compose.ui.unit.dp
+
+/**
+ * The controls for whatever settings a provider says it takes.
+ *
+ * One composable for both screens that offer them — the spawn form and the session settings dialog
+ * — and for every provider, because the server declares the list (see `DriverKind::params`) rather
+ * than this file knowing it. A driver that grows a setting gets a control here with no change to
+ * the app, which is the whole point: the values that suit one machine ship as defaults, and every
+ * one of them stays reachable from a phone.
+ *
+ * [values] is the whole map and [onChange] hands back the whole map. A key absent from it means the
+ * setting is unset, which is what every [ParamSpec.unset] describes — so clearing a field and never
+ * touching it are deliberately the same state.
+ */
+@Composable
+fun ProviderParamFields(
+    specs: List<ParamSpec>,
+    values: Map<String, String>,
+    onChange: (Map<String, String>) -> Unit,
+    /**
+     * Whether to say which settings wait for a restart. False on a spawn form, where nothing is
+     * running yet and every setting is about to be read — saying it there would be a warning about
+     * a state the reader cannot be in.
+     */
+    warnAboutRestart: Boolean,
+    modifier: Modifier = Modifier,
+) {
+    if (specs.isEmpty()) return
+    Column(modifier.fillMaxWidth()) {
+        specs.forEach { spec ->
+            val set = { value: String ->
+                onChange(
+                    // Blank clears rather than storing an empty string: the server reads an absent
+                    // key as "use the default", and an empty one would be a value it then failed
+                    // to parse.
+                    if (value.isBlank()) values - spec.key else values + (spec.key to value)
+                )
+            }
+            when (spec.kind) {
+                "choice" -> {
+                    // The first option is what unset means, so selecting it clears the key — see
+                    // `ParamKind::Choice`. Without that the picker could show a default it could
+                    // not return to.
+                    val default = spec.options.firstOrNull().orEmpty()
+                    ChipGroup(
+                        label = spec.label + restartSuffix(spec, warnAboutRestart),
+                        options = spec.options,
+                        selected = values[spec.key] ?: default,
+                        onSelect = { chosen -> set(if (chosen == default) "" else chosen) },
+                    )
+                }
+                else ->
+                    OutlinedTextField(
+                        value = values[spec.key].orEmpty(),
+                        onValueChange = set,
+                        label = { Text(spec.label + restartSuffix(spec, warnAboutRestart)) },
+                        placeholder = { Text(spec.unset) },
+                        singleLine = true,
+                        keyboardOptions = KeyboardOptions(keyboardType = keyboardFor(spec.kind)),
+                        modifier = Modifier.fillMaxWidth(),
+                    )
+            }
+            Spacer(Modifier.height(16.dp))
+        }
+        if (warnAboutRestart && specs.any { it.restart }) {
+            Text(
+                "A setting marked “on restart” is saved now and read when this session's process " +
+                    "next starts.",
+                style = MaterialTheme.typography.bodySmall,
+                color = MaterialTheme.colorScheme.onSurfaceVariant,
+            )
+        }
+    }
+}
+
+/**
+ * Marks a control whose value will not take effect yet.
+ *
+ * On the label rather than beside it, because the reader decides whether to change the thing before
+ * they touch it — a note underneath is read after the decision.
+ */
+private fun restartSuffix(spec: ParamSpec, warn: Boolean): String =
+    if (warn && spec.restart) " (on restart)" else ""
+
+/**
+ * The keyboard for a value's shape. A number field that opens the letter keyboard is one every
+ * entry is made harder by, and these are nearly all numbers.
+ */
+private fun keyboardFor(kind: String): KeyboardType =
+    when (kind) {
+        "integer" -> KeyboardType.Number
+        "decimal" -> KeyboardType.Decimal
+        else -> KeyboardType.Text
+    }
+
+/**
+ * How long typing has to stop before edited settings are sent.
+ *
+ * Long enough that a number is one request rather than one per digit, short enough that closing the
+ * dialog straight after typing still saves — the save runs on the screen behind it, which outlives
+ * the dialog, so this delay is not a window the value can be lost in.
+ */
+const val PARAM_SAVE_DELAY_MS = 700L

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

@@ -233,6 +233,8 @@ fun SessionScreen(
     // that was measured.
     var contextTokens by
         remember(address) { mutableStateOf(if (isSubagent) null else summary.contextTokens) }
+    var contextLimit by
+        remember(address) { mutableStateOf(if (isSubagent) null else summary.contextLimit) }
     // When the current compaction started. The moment comes off the `compacting` status event
     // itself -- the server timestamps every transcript line -- rather than off this device noticing
     // one, which is what makes it survive leaving the session and reopening it.
@@ -304,6 +306,14 @@ fun SessionScreen(
     // hardcoded list is a claim about a machine.
     var offeredModels by remember { mutableStateOf<List<OfferedModel>>(emptyList()) }
     var offeredPermissionModes by remember { mutableStateOf<List<String>>(emptyList()) }
+    // The settings this session's provider takes, and what they are set to. The specs come from
+    // the provider and the values from the session, because "what can be set" and "what is set"
+    // are different questions with different answers.
+    var paramSpecs by remember { mutableStateOf<List<ParamSpec>>(emptyList()) }
+    var params by remember(summary.id) { mutableStateOf(summary.params) }
+    // What was last successfully saved, so the debounce below knows whether there is anything to
+    // send -- and so a failed save can put the controls back to what the server actually holds.
+    var savedParams by remember(summary.id) { mutableStateOf(summary.params) }
     val lifecycleOwner = LocalLifecycleOwner.current
     // The resume cursor, written from the stream's IO thread.
     val lastSeq = remember { AtomicLong(0) }
@@ -489,6 +499,7 @@ fun SessionScreen(
         // Before the rest, and for every event rather than only the usage ones: a compaction and a
         // clear move this as much as a turn does. See `contextAfter`.
         contextTokens = contextAfter(contextTokens, entry.event)
+        contextLimit = contextLimitAfter(contextLimit, entry.event)
         when (val event = entry.event) {
             // Nothing further: what it carries was folded into the context above.
             is SessionEvent.UsageDelta -> {}
@@ -1205,6 +1216,30 @@ fun SessionScreen(
     fun label(id: String?): String =
         offeredModels.firstOrNull { it.id == id }?.label ?: modelLabel(id)
 
+    /**
+     * Saves edited provider settings once the typing stops.
+     *
+     * Debounced rather than sent per keystroke, because a text field over the tunnel would be a
+     * round trip and a config write per character. Here rather than in the settings dialog for the
+     * reason the delay exists at all: the dialog can be dismissed mid-edit, and this screen
+     * outlives it, so the last value typed is still saved.
+     */
+    LaunchedEffect(params) {
+        if (params == savedParams) return@LaunchedEffect
+        delay(PARAM_SAVE_DELAY_MS)
+        val wanted = params
+        try {
+            withContext(Dispatchers.IO) { setSessionParams(settings, summary.id, wanted) }
+            savedParams = wanted
+            actionError = null
+        } catch (e: ApiException) {
+            // Back to what the server holds. A control left showing a value that was refused is
+            // stating something untrue about the session.
+            params = savedParams
+            actionError = e.message
+        }
+    }
+
     // Only for the model picker, which a subagent does not have.
     if (!isSubagent) {
         LaunchedEffect(summary.machine, summary.provider) {
@@ -1218,6 +1253,7 @@ fun SessionScreen(
             }
                 .getOrNull()
             offeredPermissionModes = provider?.permissionModes.orEmpty()
+            paramSpecs = provider?.params.orEmpty()
             offeredModels =
                 provider
                     ?.let {
@@ -1916,6 +1952,7 @@ fun SessionScreen(
                 status = status,
                 compactingFor = compactingFor,
                 contextTokens = contextTokens,
+                contextLimit = contextLimit,
                 backgroundTasks = backgroundTasks,
                 subagent = isSubagent,
             )
@@ -2197,6 +2234,9 @@ fun SessionScreen(
             effort = effort.takeIf { summary.takesEffort },
             takesEffort = summary.takesEffort,
             onEffortChanged = { effort = it },
+            paramSpecs = paramSpecs,
+            params = params,
+            onParamsChanged = { params = it },
             cachedBytes = cachedBytes,
             // The purge finishes before the epoch moves, because the relaunched opening effect
             // reads the same directory and would otherwise draw what is about to be deleted. The
@@ -2473,6 +2513,8 @@ private fun SessionStatusRow(
     compactingFor: Long?,
     /** Context the session is holding, or null where nothing has measured it. */
     contextTokens: Long?,
+    /** What that is out of, or null where the provider does not say. */
+    contextLimit: Long?,
     /** Provider-reported live background work; zero is deliberately not drawn. */
     backgroundTasks: Int,
     modifier: Modifier = Modifier,
@@ -2577,7 +2619,7 @@ private fun SessionStatusRow(
         // reports usage, and one that has not run a turn all showed nothing at all, which reads as
         // a conversation with room to spare.
         Text(
-            contextTokens?.let { "context ${tokens(it)}" } ?: "context unknown",
+            contextLabel(contextTokens, contextLimit),
             style = MaterialTheme.typography.labelSmall,
             color = MaterialTheme.colorScheme.onSurfaceVariant,
         )

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

@@ -74,6 +74,15 @@ fun SessionSettingsDialog(
     onEffortChanged: (String?) -> Unit,
     /** Whether a level does anything here; the row is left out entirely where it does not. */
     takesEffort: Boolean,
+    /**
+     * The settings this session's provider takes, and what they are set to.
+     *
+     * Declared by the server rather than listed here -- see [ProviderParamFields]. Empty for a
+     * provider with none, which draws no section at all.
+     */
+    paramSpecs: List<ParamSpec>,
+    params: Map<String, String>,
+    onParamsChanged: (Map<String, String>) -> Unit,
     /**
      * What this phone is holding of the conversation, or null while that is being measured -- see
      * the Reload row below, which is what would discard it.
@@ -451,6 +460,23 @@ fun SessionSettingsDialog(
                         )
                     }
                 }
+                if (paramSpecs.isNotEmpty()) {
+                    Spacer(Modifier.height(16.dp))
+                    Text(
+                        "Model settings",
+                        style = MaterialTheme.typography.titleSmall,
+                    )
+                    Spacer(Modifier.height(8.dp))
+                    // Edited here and saved by the screen behind this, which is what makes
+                    // typing in a text field affordable: the save is debounced, and a dialog
+                    // dismissed mid-edit would take an unsaved value with it.
+                    ProviderParamFields(
+                        specs = paramSpecs,
+                        values = params,
+                        onChange = onParamsChanged,
+                        warnAboutRestart = true,
+                    )
+                }
                 Spacer(Modifier.height(8.dp))
                 Row(
                     verticalAlignment = Alignment.CenterVertically,

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

@@ -73,13 +73,9 @@ 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) }
-    var contextSize by remember { mutableStateOf("") }
+    // Whatever the chosen provider says it takes, by key. Empty until something is typed: an
-    var temperature by remember { mutableStateOf("") }
+    // absent key means the server's own default, which is what every field's placeholder says.
-    // Whether a model that carries a multi-token-prediction head drafts with it. Left to the
+    var params by remember { mutableStateOf<Map<String, String>>(emptyMap()) }
-    // 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
@@ -145,6 +141,9 @@ fun SpawnScreen(
         // machine looks exactly like one from this one.
         LaunchedEffect(machine?.id, current?.name) {
             model = ""
+            // A key from the previous provider would be a setting this one does not have, drawn
+            // by no control and sent at the spawn anyway.
+            params = emptyMap()
             providerModels = emptyList()
             providerModelsError = null
             permissionMode = current?.defaultPermissionMode.orEmpty()
@@ -283,35 +282,14 @@ fun SpawnScreen(
             Spacer(Modifier.height(16.dp))
         }
 
-        if (isLlama) {
+        // Nothing is running yet, so nothing here waits for a restart -- every one of these is
-            OutlinedTextField(
+        // read by the process this form is about to start.
-                value = contextSize,
+        ProviderParamFields(
-                onValueChange = { contextSize = it },
+            specs = current?.params.orEmpty(),
-                label = { Text("Context size (blank = the model's default)") },
+            values = params,
-                singleLine = true,
+            onChange = { params = it },
-                modifier = Modifier.fillMaxWidth(),
+            warnAboutRestart = false,
-            )
+        )
-            Spacer(Modifier.height(16.dp))
-
-            OutlinedTextField(
-                value = temperature,
-                onValueChange = { temperature = it },
-                label = { Text("Temperature (blank = llama.cpp's default)") },
-                singleLine = true,
-                modifier = Modifier.fillMaxWidth(),
-            )
-            Spacer(Modifier.height(16.dp))
-
-            // 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.
@@ -387,28 +365,10 @@ fun SpawnScreen(
                                     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
+                                    // Already only the keys somebody set: a field left blank
-                                    // by default" rather than a zero.
+                                    // removes its key rather than sending an empty value, so
-                                    params =
+                                    // "blank" reaches the server as "your default".
-                                        buildMap {
+                                    params = params,
-                                            if (isLlama) {
-                                                contextSize
-                                                    .trim()
-                                                    .takeIf { it.isNotEmpty() }
-                                                    ?.let { put("contextSize", it) }
-                                                temperature
-                                                    .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")
-                                                }
-                                            }
-                                        },
                                 )
                             }
                         onSpawned(spawned)
@@ -426,15 +386,6 @@ fun SpawnScreen(
     }
 }
 
-/**
- * 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.
  *

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

@@ -555,6 +555,7 @@ fun foldEvent(items: List<TranscriptItem>, entry: SeqEvent): List<TranscriptItem
         is SessionEvent.RetiredTaskNote -> items
         // Screen-level state, not transcript rows -- see SessionScreen.
         is SessionEvent.UsageDelta -> items
+        is SessionEvent.ContextWindow -> items
     }
 
 /**

app/androidApp/src/test/kotlin/com/example/aiapp/SessionOrderTest.kt

@@ -75,6 +75,8 @@ class SessionOrderTest {
             resumeAt = null,
             cwd = null,
             contextTokens = null,
+            contextLimit = null,
+            params = emptyMap(),
             maxImageEdge = null,
             usageProvider = null,
             status = status,

server/src/config.rs

@@ -312,6 +312,28 @@ impl DriverKind {
         }
     }
 
+    /// The settings this kind of session takes beyond the shared ones, for
+    /// the phone to offer.
+    ///
+    /// Declared rather than drawn: a spawn screen with a field per llama
+    /// setting is a screen that has to be edited every time a driver grows
+    /// one, and this app already has the `params` map to carry them. So the
+    /// server says what a provider takes and the phone renders it, which is
+    /// the same arrangement `permission_modes` uses and for the same reason
+    /// -- the alternative is two lists that disagree, one of them in Kotlin.
+    ///
+    /// It is also what keeps these *reachable at all*. Several were hardcoded
+    /// to the values measured on one machine, which is fine as a default and
+    /// wrong as a constant: the next machine has a different GPU and a
+    /// different number of cores, and nobody running this app can edit the
+    /// source.
+    pub fn params(self) -> &'static [ParamSpec] {
+        match self {
+            Self::LlamaCpp => LLAMA_PARAMS,
+            Self::Echo | Self::ClaudeCli | Self::CodexCli => &[],
+        }
+    }
+
     /// The mode used when a new-session form first selects this kind.
     pub fn default_permission_mode(self) -> Option<&'static str> {
         match self {
@@ -323,6 +345,134 @@ impl DriverKind {
     }
 }
 
+/// One setting a provider takes, and enough about it to draw a control.
+///
+/// Deliberately thin: a key, words for a person, and which shape the value
+/// has. Anything richer -- units, validation, dependencies between settings --
+/// would be a schema language, and what the phone needs is a text field or a
+/// row of chips.
+#[derive(Debug, Clone, Copy, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct ParamSpec {
+    /// The `SessionConfig::params` key this writes.
+    pub key: &'static str,
+    pub label: &'static str,
+    /// What happens when it is not set, in words. Shown where a control shows
+    /// its placeholder, so "blank" always means something specific rather than
+    /// leaving the reader to guess whether it means zero.
+    pub unset: &'static str,
+    /// Flattened, so a spec is one flat object: `kind` beside the rest rather
+    /// than an object of its own with `kind` inside it.
+    #[serde(flatten)]
+    pub kind: ParamKind,
+    /// Whether changing it waits for the process to start again.
+    ///
+    /// The honest half of offering these live. A sampling setting rides on the
+    /// next request; a server flag was decided when the model was loaded, and
+    /// a control that silently did nothing until some later restart would be
+    /// worse than one that is not there.
+    pub restart: bool,
+}
+
+/// What shape a [`ParamSpec`]'s value has.
+#[derive(Debug, Clone, Copy, Serialize)]
+#[serde(rename_all = "camelCase", tag = "kind")]
+pub enum ParamKind {
+    Integer,
+    Decimal,
+    Text,
+    /// A fixed set. **The first option is what leaving it unset means**, and
+    /// choosing it clears the setting rather than storing a value -- so the
+    /// default is a state the picker can return to, and the stored config
+    /// does not fill up with values nobody chose.
+    Choice {
+        options: &'static [&'static str],
+    },
+}
+
+/// What a llama.cpp session takes.
+///
+/// The server flags first, in the order they matter, then the sampling ones --
+/// which is also the order of how disruptive changing one is.
+const LLAMA_PARAMS: &[ParamSpec] = &[
+    ParamSpec {
+        key: "contextSize",
+        label: "Context size",
+        unset: "the model's own trained context",
+        kind: ParamKind::Integer,
+        restart: true,
+    },
+    ParamSpec {
+        key: "tools",
+        label: "Tools",
+        // Worth a control rather than a constant because of what it costs:
+        // the definitions of all seven are ~2,000 tokens of the context,
+        // every turn, before anything is said. On a small window that is the
+        // difference between a usable session and one that overruns.
+        unset: "all of them -- or a comma-separated list, or \"none\"",
+        kind: ParamKind::Text,
+        restart: true,
+    },
+    ParamSpec {
+        key: "gpuLayers",
+        label: "Layers on the GPU",
+        unset: "as many as fit",
+        kind: ParamKind::Integer,
+        restart: true,
+    },
+    ParamSpec {
+        key: "threads",
+        label: "Threads",
+        unset: "one per core",
+        kind: ParamKind::Integer,
+        restart: true,
+    },
+    ParamSpec {
+        key: "speculative",
+        label: "Speculative decoding",
+        unset: "on, for a model whose file carries a draft head",
+        kind: ParamKind::Choice {
+            options: &["auto", "off"],
+        },
+        restart: true,
+    },
+    ParamSpec {
+        key: "specDraftNMax",
+        label: "Tokens drafted ahead",
+        unset: "llama.cpp's own default",
+        kind: ParamKind::Integer,
+        restart: true,
+    },
+    ParamSpec {
+        key: "temperature",
+        label: "Temperature",
+        unset: "llama.cpp's default",
+        kind: ParamKind::Decimal,
+        restart: false,
+    },
+    ParamSpec {
+        key: "topP",
+        label: "Top P",
+        unset: "llama.cpp's default",
+        kind: ParamKind::Decimal,
+        restart: false,
+    },
+    ParamSpec {
+        key: "topK",
+        label: "Top K",
+        unset: "llama.cpp's default",
+        kind: ParamKind::Integer,
+        restart: false,
+    },
+    ParamSpec {
+        key: "maxTokens",
+        label: "Reply limit",
+        unset: "no limit",
+        kind: ParamKind::Integer,
+        restart: false,
+    },
+];
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub struct TokenEntry {

server/src/routes.rs

@@ -54,6 +54,8 @@
 //!                                     which starts again in the new one
 //! POST   /sessions/{id}/model          {model}
 //! POST   /sessions/{id}/permission-mode  {permissionMode}
+//! POST   /sessions/{id}/params         {params} -- the provider settings, whole;
+//!                                      what a provider takes is on its ProviderInfo
 //! POST   /sessions/{id}/effort         {effort} -- null for the CLI's default;
 //!                                     settled at launch, so this stops the process
 //! POST   /sessions/{id}/command        {text}  -- /compact, /clear, /rename x, or the dialect's own
@@ -166,6 +168,7 @@ pub fn router(manager: Arc<SessionManager>) -> Router {
         .route("/sessions/{id}/cwd", post(set_cwd))
         .route("/sessions/{id}/model", post(set_model))
         .route("/sessions/{id}/permission-mode", post(set_permission_mode))
+        .route("/sessions/{id}/params", post(set_params))
         .route("/sessions/{id}/effort", post(set_effort))
         .route("/defaults", get(defaults).post(set_defaults))
         .route("/sessions/{id}/notify", post(set_notify))
@@ -300,6 +303,11 @@ struct ProviderInfo {
     permission_modes: Vec<&'static str>,
     #[serde(skip_serializing_if = "Option::is_none")]
     default_permission_mode: Option<&'static str>,
+    /// The per-kind settings a spawn form and the session settings dialog
+    /// offer -- see [`crate::config::DriverKind::params`]. Empty for a
+    /// provider with none, which draws no section at all.
+    #[serde(skip_serializing_if = "<[_]>::is_empty")]
+    params: &'static [crate::config::ParamSpec],
 }
 
 async fn list_machines(State(manager): State<Arc<SessionManager>>) -> axum::Json<Vec<MachineInfo>> {
@@ -320,6 +328,7 @@ fn info_for(machine: crate::config::MachineConfig) -> MachineInfo {
                 models: provider.models,
                 permission_modes: provider.kind.permission_modes().to_vec(),
                 default_permission_mode: provider.kind.default_permission_mode(),
+                params: provider.kind.params(),
             })
             .collect(),
     }
@@ -426,6 +435,7 @@ async fn probe_machine(
                 models: provider.models,
                 permission_modes: provider.kind.permission_modes().to_vec(),
                 default_permission_mode: provider.kind.default_permission_mode(),
+                params: provider.kind.params(),
             })
             .collect(),
     ))
@@ -1586,6 +1596,24 @@ async fn set_permission_mode(
     Ok(StatusCode::NO_CONTENT)
 }
 
+#[derive(Deserialize)]
+#[serde(deny_unknown_fields)]
+struct ParamsRequest {
+    /// The whole map, not a patch -- see `SessionManager::set_session_params`.
+    params: std::collections::BTreeMap<String, String>,
+}
+
+async fn set_params(
+    State(manager): State<Arc<SessionManager>>,
+    UrlPath(id): UrlPath<String>,
+    axum::Json(body): axum::Json<ParamsRequest>,
+) -> Result<StatusCode, ApiError> {
+    manager
+        .set_session_params(&id, body.params)
+        .map_err(bad_request)?;
+    Ok(StatusCode::NO_CONTENT)
+}
+
 #[derive(Deserialize)]
 #[serde(deny_unknown_fields)]
 struct NotifyRequest {

server/src/session/driver.rs

@@ -363,6 +363,26 @@ pub enum Event {
         #[serde(default, skip_serializing_if = "Option::is_none")]
         permission_mode: Option<String>,
     },
+    /// How much context this session's model has to hold a conversation in.
+    ///
+    /// The denominator the phone draws [`Event::UsageDelta`]'s `context`
+    /// against, and its own event rather than a field on that one because it
+    /// is not a per-turn measurement: it is fixed when the process starts and
+    /// changes only when a different one is started, which is what a model
+    /// change does. Reported the moment it is known, so the figure and what it
+    /// is out of arrive together rather than the first turn drawing a
+    /// numerator with no denominator.
+    ///
+    /// **Only ever sent by a driver that actually knows.** A window nobody has
+    /// measured is not an unlimited one: llama.cpp answers it exactly, because
+    /// the number is a flag the server was started with and `/props` reads it
+    /// back, while a coding CLI's context is the vendor's business and
+    /// nothing in either control protocol states it. Those send nothing, the
+    /// session has no limit, and the phone draws the figure on its own -- see
+    /// `SessionSummary::context_limit`.
+    ContextWindow {
+        tokens: u64,
+    },
     /// Per-turn token counts, where the dialect reports them.
     UsageDelta {
         /// What this turn cost: the tokens it was charged for.
@@ -494,6 +514,25 @@ pub fn context_tokens(input: u64, cache_creation: u64, cache_read: u64) -> u64 {
     input + cache_creation + cache_read
 }
 
+/// The context window after `event`, given what it was before.
+///
+/// Beside [`context_after`] because it has the same three readers and the same
+/// hazard: a figure that outlives what made it true. A model change replaces
+/// the process, so the window it reports replaces the old one -- and until the
+/// new one says, there is no answer rather than the previous model's.
+pub fn context_limit_after(current: Option<u64>, event: &Event) -> Option<u64> {
+    match event {
+        Event::ContextWindow { tokens } => Some(*tokens),
+        // The window belongs to the process, and a stopped one has none. Left
+        // standing, a restarted session on a different model would draw its
+        // occupancy against the previous model's window.
+        Event::Status {
+            state: SessionStatus::Exited,
+        } => None,
+        _ => current,
+    }
+}
+
 /// The context after `event`, given what it was before.
 ///
 /// The whole rule in one place, because three readers need the same answer:
@@ -683,6 +722,16 @@ pub trait Driver: Send + Sync {
     /// `/rename` afterwards, which is what puts the same name in its own
     /// session picker and in what other agents see.
     fn set_title(&self, title: &str);
+    /// Takes the session's provider settings, whole.
+    ///
+    /// The whole map because it is a form's contents -- an absent key means
+    /// "unset", not "unchanged". A driver applies what it can apply now and
+    /// says so about the rest: the map is also on disk by the time this is
+    /// called, so a setting that only takes effect at the next start is not
+    /// lost, it is waiting. The default is right for a driver with no settings
+    /// of its own, which is every one but llama.cpp -- see
+    /// [`crate::config::DriverKind::params`].
+    fn set_params(&self, _params: &std::collections::BTreeMap<String, String>) {}
     /// Runs a command this session's own dialect understands, verbatim --
     /// `/context`, `/usage`, anything a CLI adds next month. A driver with no
     /// such vocabulary says so with an [`Event::Error`] rather than sending it

server/src/session/llama/mod.rs

@@ -65,6 +65,32 @@ pub use tools::{DEFAULT_MODE as DEFAULT_PERMISSION_MODE, MODES as PERMISSION_MOD
 use mcp::McpServer;
 use tools::Tools;
 
+/// The sampling half of a session's settings, in the wire's own names.
+///
+/// One function because two callers need the identical mapping: the spawn, and
+/// a later change through [`Driver::set_params`]. A value that will not parse
+/// as a number is left out rather than passed through -- `llama-server` would
+/// refuse the whole request for it, which would look like the session breaking
+/// rather than like one field being wrong.
+fn sampling_from(
+    params: &std::collections::BTreeMap<String, String>,
+) -> serde_json::Map<String, Value> {
+    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) = params.get(key)
+            && let Ok(number) = raw.parse::<f64>()
+        {
+            sampling.insert(field.to_string(), json!(number));
+        }
+    }
+    sampling
+}
+
 /// Exa's own hosted MCP server, which is what a llama session searches the web
 /// with. The address llama.cpp's web UI offers under "Exa" in its recommended
 /// servers, so a session here reaches the same thing that UI does.
@@ -76,6 +102,11 @@ pub const EXA_MCP_URL: &str = "https://mcp.exa.ai/mcp";
 /// is the escape for a machine where drafting turns out not to pay.
 const SPECULATIVE: &str = "speculative";
 
+/// The spawn parameter naming which built-in tools a session gets, as
+/// `llama-server`'s own comma-separated list. Absent is all of them, and
+/// `"none"` is the way to ask for a session that only talks.
+const TOOLS: &str = "tools";
+
 /// How many times one message may go round the call-a-tool loop.
 ///
 /// A bound rather than a budget: a small model that has decided to read the
@@ -221,8 +252,10 @@ struct Shared {
     /// Where this session's process record lives, so [`Driver::stop`] can find
     /// the server it has to end.
     session_dir: PathBuf,
-    /// Sampling settings chosen at spawn, sent with every request.
+    /// Sampling settings, sent with every request. Behind a lock because they
-    sampling: serde_json::Map<String, Value>,
+    /// are changeable while the session runs: they ride on the next request,
+    /// so unlike the server's own flags there is nothing to reload.
+    sampling: Mutex<serde_json::Map<String, Value>>,
     /// The session's working directory, which is where its tools act. `None`
     /// leaves that to `llama-server`, which is the honest answer rather than a
     /// guess at one.
@@ -289,26 +322,14 @@ impl LlamaDriver {
             "a llama.cpp session needs a model -- one of the downloaded ones, by its key",
         )?;
 
-        let mut sampling = serde_json::Map::new();
+        let sampling = sampling_from(&meta.params);
-        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));
-            }
-        }
 
         let driver = Self {
             shared: Arc::new(Shared {
                 sink,
                 transcript: transcript.to_path_buf(),
                 session_dir: session_dir.to_path_buf(),
-                sampling,
+                sampling: Mutex::new(sampling),
                 cwd: meta
                     .cwd
                     .as_ref()
@@ -403,6 +424,15 @@ impl LlamaDriver {
                         "{model} loaded and answering at {endpoint} with {} tools",
                         tools.offered().map_or(0, |offered| offered.len()),
                     );
+                    // Asked now rather than carried from the spawn flags: a
+                    // session that named no context size gets the model's
+                    // own, which only the loaded server knows, and one that
+                    // named an impossible one gets whatever it settled for.
+                    // Either way this is the measurement rather than the
+                    // request.
+                    if let Some(window) = context_window(&endpoint) {
+                        shared.emit(Event::ContextWindow { tokens: window });
+                    }
                     Serving::Ready {
                         endpoint: endpoint.clone(),
                         tools: Arc::new(tools),
@@ -531,14 +561,12 @@ fn spawn_server(
         "127.0.0.1".into(),
         "--port".into(),
         forward.there.to_string(),
-        // The built-in agent tools -- read, search, edit, shell. Every one of
+        // The built-in agent tools -- read, search, edit, shell. All of them
-        // them, because a session offered a subset is a session that says "I
+        // unless the session says otherwise, because whether a particular call
-        // can't do that" about something it was installed to do, and the
+        // should happen is the permission gate's question rather than a flag's.
-        // question of whether a particular call should happen is the
+        // They run on the machine serving the model, which is the machine the
-        // permission gate's rather than a flag's. They run on the machine
+        // files are on.
-        // serving the model, which is the machine the files are on.
+        //
-        "--tools".into(),
-        "all".into(),
         // One slot, not the four `llama-server` picks on its own. A session is
         // one conversation making one request at a time -- the driver holds a
         // second message until the turn ends -- so the other three are context
@@ -553,6 +581,27 @@ fn spawn_server(
         "-np".into(),
         "1".into(),
     ];
+    // Settable because it is not free: the definitions of all seven are around
+    // 2,000 tokens of every prompt -- measured at 2,191 against 1,322 for two
+    // of them -- which on a small context window is a quarter of it spent
+    // before anything is said.
+    //
+    // "none" omits the flag rather than passing it on: `--tools none` is
+    // `tools setup failed: unknown tool "none"` and a server that exits, since
+    // the argument is a list of tool names and no-tools is what having no flag
+    // means.
+    match meta.params.get(TOOLS).map(|chosen| chosen.trim()) {
+        Some("none") => {}
+        chosen => {
+            args.push("--tools".into());
+            args.push(
+                chosen
+                    .filter(|c| !c.is_empty())
+                    .unwrap_or("all")
+                    .to_string(),
+            );
+        }
+    }
     // A model that carries a multi-token-prediction head drafts with it, which
     // is most of a 50% speed-up for free -- the tensors are in the file
     // whether or not they are used, and without the flag `llama-server` says
@@ -573,6 +622,11 @@ fn spawn_server(
         ("contextSize", "-c"),
         ("gpuLayers", "-ngl"),
         ("threads", "-t"),
+        // How far ahead the draft head guesses. Not defaulted here: 2 measured
+        // 7% faster than llama.cpp's 3 on this machine's GPU, once, which is
+        // a reason to make the knob reachable and not a reason to move it for
+        // everybody.
+        ("specDraftNMax", "--spec-draft-n-max"),
     ] {
         if let Some(value) = meta.params.get(key) {
             args.push(flag.to_string());
@@ -854,7 +908,10 @@ fn converse(
         if shared.cancel.load(Ordering::SeqCst) {
             return Ok(());
         }
-        let reply = generate(endpoint, &messages, tools, &shared.sampling, shared)?;
+        // Read per call rather than per turn, so a sampling change made while
+        // a long turn is running reaches the rest of it.
+        let sampling = shared.sampling.lock().unwrap().clone();
+        let reply = generate(endpoint, &messages, tools, &sampling, shared)?;
         let calls = reply.calls;
         messages.push(Message {
             tool_calls: calls.iter().map(Call::wire).collect(),
@@ -1046,6 +1103,49 @@ impl Driver for LlamaDriver {
     // is called, and the rename has already happened where the name lives.
     fn set_title(&self, _title: &str) {}
 
+    /// Takes new settings: the sampling half now, and says so about the rest.
+    ///
+    /// The split is what [`crate::config::ParamSpec::restart`] describes, and
+    /// it is said out loud rather than left to the screen, because the screen
+    /// can only say what a setting *usually* does -- this is the one place
+    /// that knows whether this session's server was started with the old
+    /// value. A session already stopped needs no such note: its next start
+    /// will read all of them.
+    fn set_params(&self, params: &std::collections::BTreeMap<String, String>) {
+        *self.shared.sampling.lock().unwrap() = sampling_from(params);
+        // Only the settings that actually differ from what this session's
+        // server was started with. Listing every restart-only one on every
+        // save would be a wall of text about nothing having changed.
+        let waiting: Vec<&str> = crate::config::DriverKind::LlamaCpp
+            .params()
+            .iter()
+            .filter(|spec| {
+                spec.restart && params.get(spec.key) != self.respawn.meta.params.get(spec.key)
+            })
+            .map(|spec| spec.label)
+            .collect();
+        // Nothing to say to a session with no server: its next start reads all
+        // of them, which is what the note would have been asking for.
+        let running = matches!(&*self.shared.serving.lock().unwrap(), Serving::Ready { .. });
+        if !waiting.is_empty() && running {
+            let one = waiting.len() == 1;
+            self.shared.emit(Event::Error {
+                message: format!(
+                    "{} {} saved. {} when this session's server next starts -- stop and start \
+                     the session, or change its model, to load {} now.",
+                    waiting.join(", "),
+                    if one { "is" } else { "are" },
+                    if one {
+                        "It takes effect"
+                    } else {
+                        "They take effect"
+                    },
+                    if one { "it" } else { "them" },
+                ),
+            });
+        }
+    }
+
     fn set_permission_mode(&self, mode: &str) {
         if !tools::MODES.contains(&mode) {
             self.shared.emit(Event::Error {
@@ -1436,6 +1536,23 @@ fn log_tail(session_dir: &Path) -> String {
 /// How much of that log to carry into a message somebody reads on a phone.
 const LOG_TAIL_LINES: usize = 6;
 
+/// How many tokens this server can hold, from the server itself.
+///
+/// `/props` reports the per-slot context, which is the whole of it because
+/// this driver always starts one slot -- see the `-np` argument. `None` for a
+/// server that would not answer, which draws as no ceiling rather than as a
+/// guessed one.
+fn context_window(endpoint: &str) -> Option<u64> {
+    ureq::get(format!("{endpoint}/props"))
+        .call()
+        .ok()?
+        .body_mut()
+        .read_json::<Value>()
+        .ok()?
+        .pointer("/default_generation_settings/n_ctx")
+        .and_then(Value::as_u64)
+}
+
 /// One streamed completion: posts the conversation, emits each text delta as
 /// it arrives, and assembles whatever tool calls came with it.
 ///
@@ -1491,9 +1608,16 @@ fn generate(
     // at once interleaves their fragments, each tagged with its index.
     let mut calls: Vec<Call> = Vec::new();
     let mut tokens = 0u64;
-    // The prompt side only, which is what the model is holding -- the same
+    // What the model is holding when this call ends: the prompt it was given
-    // definition the other dialects report, so one word on the phone means one
+    // plus the reply it produced, which is exactly what the next call's prompt
-    // thing whichever kind of session it is.
+    // begins with.
+    //
+    // The prompt alone was wrong, and measurably: a five-word reply reported
+    // 2,042 against a slot holding 2,355 (2026-09-19, checked against
+    // `stop processing: n_tokens` in the server's own log). The gap is the
+    // reply, so it grows with how much the model just said -- which is the
+    // worst direction for a figure somebody is watching to see how much room
+    // is left.
     let mut context = None;
     for line in std::io::BufRead::lines(reader) {
         if shared.cancel.load(Ordering::SeqCst) {
@@ -1510,13 +1634,13 @@ fn generate(
         let Ok(chunk) = serde_json::from_str::<Value>(payload) else {
             continue;
         };
-        if let Some(usage) = chunk.get("usage") {
+        // One figure answering both questions, which for this dialect it
-            if let Some(total) = usage.get("total_tokens").and_then(Value::as_u64) {
+        // genuinely does: what the call was charged for and what the model is
-                tokens = total;
+        // left holding are the same tokens, because nothing here is billed and
-            }
+        // the whole conversation is resent every time.
-            if let Some(prompt) = usage.get("prompt_tokens").and_then(Value::as_u64) {
+        if let Some(total) = chunk.pointer("/usage/total_tokens").and_then(Value::as_u64) {
-                context = Some(prompt);
+            tokens = total;
-            }
+            context = Some(total);
         }
         let Some(delta) = chunk.pointer("/choices/0/delta") else {
             continue;

server/src/session/llama/tools.rs

@@ -57,18 +57,29 @@ 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
+    /// A session with no built-in tools -- or none at all -- is a perfectly
-    /// session with only MCP tools is a perfectly good session -- so nothing
+    /// good session, so nothing here treats "no tools" as a failure. What *is*
-    /// here treats "no tools" as a failure. What *is* a failure is not being
+    /// a failure is not being able to ask at all, because that is the same
-    /// able to ask, because that is the same server the conversation is about
+    /// server the conversation is about to go to.
-    /// to go to.
     pub fn discover(endpoint: &str, mcp: Vec<Arc<Mutex<McpServer>>>) -> Result<Self> {
-        let catalog: Vec<Value> = ureq::get(format!("{endpoint}/tools"))
+        let mut response = ureq::get(format!("{endpoint}/tools"))
+            // A server started with no `--tools` answers **403** here, not an
+            // empty list -- the route is off rather than empty. Read as a
+            // failure that was a session which never started, for the one
+            // setting whose whole purpose is to have no tools.
+            .config()
+            .http_status_as_error(false)
+            .build()
             .call()
-            .context("asking llama-server which tools it has")?
+            .context("asking llama-server which tools it has")?;
-            .body_mut()
+        let catalog: Vec<Value> = if response.status().is_success() {
-            .read_json()
+            response
-            .context("reading llama-server's tool list")?;
+                .body_mut()
+                .read_json()
+                .context("reading llama-server's tool list")?
+        } else {
+            Vec::new()
+        };
         let mut definitions = Vec::new();
         let mut server = HashMap::new();
         for entry in &catalog {

server/src/session/mod.rs

@@ -36,7 +36,8 @@ use crate::config::{
 use claude::ClaudeDriver;
 use codex::CodexDriver;
 use driver::{
-    AttachmentRef, Driver, Event, EventSink, SessionCommand, SessionStatus, Unqueued, context_after,
+    AttachmentRef, Driver, Event, EventSink, SessionCommand, SessionStatus, Unqueued,
+    context_after, context_limit_after,
 };
 use echo::EchoDriver;
 use llama::LlamaDriver;
@@ -213,6 +214,17 @@ pub struct SessionInfo {
     /// are different answers and the phone draws them differently.
     #[serde(skip_serializing_if = "Option::is_none")]
     pub context_tokens: Option<u64>,
+    /// What that figure is out of -- see [`Event::ContextWindow`]. Absent
+    /// where the provider does not say, which is a third state again: not a
+    /// session with room to spare, and not one whose occupancy is unknown,
+    /// but one whose occupancy is known and whose ceiling is not.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub context_limit: Option<u64>,
+    /// The provider settings this session was launched with, as the settings
+    /// dialog has to open on them -- what a control is *set to* is not
+    /// derivable from what the provider *offers*.
+    #[serde(skip_serializing_if = "std::collections::BTreeMap::is_empty")]
+    pub params: std::collections::BTreeMap<String, String>,
     /// The longest edge an image should have by the time it gets here.
     /// Absent rather than a large number, because "no limit" and "a limit
     /// that happens to be big" are different answers.
@@ -415,6 +427,7 @@ struct Shared {
     /// the session row so a phone opening a long conversation has the real
     /// figure rather than whatever its newest page mentions.
     context_tokens: Mutex<Option<u64>>,
+    context_limit: Mutex<Option<u64>>,
     /// Mirrored out of the config so the pump can read it without taking
     /// the manager's lock -- the pump runs underneath the manager, and
     /// reaching back up would invert that.
@@ -579,6 +592,11 @@ impl LiveSession {
             effort: current.effort.clone(),
             takes_effort: kind.is_some_and(DriverKind::takes_effort),
             context_tokens: *self.shared.context_tokens.lock().unwrap(),
+            context_limit: *self.shared.context_limit.lock().unwrap(),
+            // From the config for the reason `effort` above is: the settings
+            // the process was started with are what a restart-only control has
+            // to show, and this is where they are kept.
+            params: current.params.clone(),
             notify: *self.shared.notify.lock().unwrap(),
             auto_resume: current.auto_resume,
             auto_resume_message: resume_message(current),
@@ -1104,6 +1122,8 @@ impl SessionManager {
                     takes_effort: kind_of(&inner.config, &meta.machine, &meta.provider)
                         .is_some_and(DriverKind::takes_effort),
                     context_tokens: None,
+                    context_limit: None,
+                    params: meta.params.clone(),
                     max_image_edge: kind_of(&inner.config, &meta.machine, &meta.provider)
                         .and_then(DriverKind::max_image_edge),
                     usage_provider: kind_of(&inner.config, &meta.machine, &meta.provider)
@@ -1319,6 +1339,43 @@ impl SessionManager {
         Ok(())
     }
 
+    /// Changes this session's provider settings, live and persisted.
+    ///
+    /// The whole map rather than one key, because that is what a settings
+    /// screen has: a form is submitted as its contents, and merging one field
+    /// at a time would make clearing a field indistinguishable from not
+    /// mentioning it. An absent key *is* the instruction to unset it.
+    ///
+    /// Persisted first for the reason the model is: the config answers what to
+    /// launch with next time, which is the whole of what a restart-only
+    /// setting means. The driver is then told, and takes what it can use now.
+    pub fn set_session_params(
+        &self,
+        id: &str,
+        params: std::collections::BTreeMap<String, String>,
+    ) -> Result<()> {
+        let mut inner = self.inner.write().unwrap();
+        if !inner.config.sessions.iter().any(|meta| meta.id == id) {
+            bail!("no session {id}");
+        }
+        let mut candidate = inner.config.clone();
+        for meta in candidate.sessions.iter_mut().filter(|meta| meta.id == id) {
+            meta.params = params.clone();
+        }
+        candidate.save(&self.config_path)?;
+        inner.config = candidate;
+        // Told to the driver where there is one, and simply saved where there
+        // is not. Deliberately not `ask`: a session with no process has not
+        // failed to take these, it has taken them in the only way that matters
+        // -- its next start reads them -- and adjusting the settings of a
+        // stopped session so that starting it uses them is the ordinary thing
+        // to do, not an error to report.
+        if let Some(driver) = inner.live.get(id).and_then(|session| session.driver()) {
+            driver.set_params(&params);
+        }
+        Ok(())
+    }
+
     /// Turns this session's notifications on or off, live and persisted --
     /// both, or the switch moves back on its own at the next restart.
     ///
@@ -2394,6 +2451,7 @@ fn launch(
         model: Mutex::new(meta.model.clone()),
         permission_mode: Mutex::new(meta.permission_mode.clone()),
         context_tokens: Mutex::new(transcript.context_tokens()),
+        context_limit: Mutex::new(transcript.context_limit()),
         notify: Mutex::new(meta.notify),
         written: Mutex::new(0),
     });
@@ -2674,6 +2732,10 @@ async fn pump(
             let mut context = shared.context_tokens.lock().unwrap();
             *context = context_after(*context, &event);
         }
+        {
+            let mut limit = shared.context_limit.lock().unwrap();
+            *limit = context_limit_after(*limit, &event);
+        }
         // Nothing changed, so there is nothing to record. Both of these
         // repeat: an imported session reads the turn state off its file's
         // newest record on every sync, and the CLI restates its model and

server/src/session/transcript.rs

@@ -15,7 +15,7 @@ use std::path::Path;
 use anyhow::{Context, Result};
 use serde::{Deserialize, Serialize};
 
-use super::driver::{Event, SessionStatus, context_after};
+use super::driver::{Event, SessionStatus, context_after, context_limit_after};
 
 /// One transcript line: an [`Event`] plus its position and time. The event
 /// is flattened so the wire shape stays one flat object.
@@ -34,6 +34,7 @@ pub struct Transcript {
     last_status: Option<SessionStatus>,
     last_activity: Option<f64>,
     context_tokens: Option<u64>,
+    context_limit: Option<u64>,
 }
 
 impl Transcript {
@@ -68,6 +69,12 @@ impl Transcript {
             context_tokens: existing
                 .iter()
                 .fold(None, |current, entry| context_after(current, &entry.event)),
+            // The same fold for the same reason: a session whose process has
+            // since exited has no window, and the newest `ContextWindow` line
+            // alone would not know that.
+            context_limit: existing.iter().fold(None, |current, entry| {
+                context_limit_after(current, &entry.event)
+            }),
         })
     }
 
@@ -107,6 +114,12 @@ impl Transcript {
         self.context_tokens
     }
 
+    /// What that figure is out of, as of opening, and `None` where this
+    /// session's provider does not say.
+    pub fn context_limit(&self) -> Option<u64> {
+        self.context_limit
+    }
+
     /// Appends `event`, assigning it the next sequence number. Flushed per
     /// event: each line is tiny, and the transcript is the source of truth a
     /// crash must not lose the tail of.