@otto-code/server 0.7.1 → 0.7.2

package/dist/server/server/agent/agent-manager.d.ts

@@ -243,6 +243,17 @@ interface ManagedAgentBase {
      * raced into a second concurrent turn.
      */
     pendingSteerDrain: boolean;
+    /**
+     * Set when a cancel (the composer's stop button, ESC, `cancel_agent`) lands
+     * while messages are queued. Stop means stop, so the cancelled turn's
+     * finalize must not drain the queue into a fresh turn — but the entries
+     * survive, so the Queue track still shows them and the user can edit or send
+     * them when ready. Cleared the moment the next run starts, after which the
+     * queue drains normally again.
+     *
+     * Optional so existing `ManagedAgent` fixtures stay valid.
+     */
+    steerQueueHeld?: boolean;
     persistence: AgentPersistenceHandle | null;
     historyPrimed: boolean;
     lastUserMessageAt: Date | null;
@@ -691,8 +702,23 @@ export declare class AgentManager {
     /** Pull one entry back out of the queue (Queue-track edit / send now). */
     removeSteerQueueEntry(agentId: string, entryId: string): SteerQueueEntry | null;
     /**
-     * Drop every queued message. Cancelling an agent is one "stop everything"
-     * verb: aborting the run also abandons the work you had lined up behind it.
+     * Keep the queue, but stop the cancelled turn from draining it.
+     *
+     * Cancel used to empty the queue outright, on the theory that "stop
+     * everything" includes the work lined up behind it. In practice the messages
+     * you queued are the ones you still want: stopping the run is how you make
+     * room for them, so wiping them destroys work the user typed. So stop now
+     * means exactly what it says — nothing new starts by itself — while the
+     * entries stay in the Queue track, ready to edit or send.
+     *
+     * Returns how many entries were held (0 when there was no queue), purely for
+     * logging.
+     */
+    holdSteerQueue(agentId: string): number;
+    /**
+     * Drop every queued message — the explicit "clear the queue" verb behind
+     * `agent.queue.clear`, and the way a closing agent sheds its queue. Cancel
+     * does NOT come through here; see `holdSteerQueue`.
      */
     clearSteerQueue(agentId: string): number;
     replaceAgentRun(agentId: string, prompt: AgentPromptInput, options?: AgentRunOptions): AsyncGenerator<AgentStreamEvent>;

package/dist/server/server/agent/agent-manager.js

@@ -1943,6 +1943,9 @@ export class AgentManager {
         const agent = existingAgent;
         agent.pendingReplacement = false;
         agent.pendingSteerDrain = false;
+        // A new run resumes the queue: the hold only ever covered the finalize of
+        // the turn that was cancelled.
+        agent.steerQueueHeld = false;
         agent.lastError = undefined;
         const pendingRun = this.foregroundRuns.createPendingRun(agentId);
         const streamForwarder = async function* streamForwarder() {
@@ -2008,8 +2011,10 @@ export class AgentManager {
         // buffered instead of racing into a second concurrent turn. A terminal
         // error (or a replacement already holding the slot) skips the drain: a
         // queued turn must never run unprompted into a broken session — the queue
-        // is held and surfaced so the supervisor decides.
-        const drainBatch = !shouldHoldBusyForReplacement && !terminalError
+        // is held and surfaced so the supervisor decides. A cancel holds it for the
+        // same reason: the user pressed stop, so nothing new starts on its own,
+        // and the queue stays put for them to send when ready.
+        const drainBatch = !shouldHoldBusyForReplacement && !terminalError && !mutableAgent.steerQueueHeld
             ? takeNextSteerQueueBatch(mutableAgent.steerQueue)
             : null;
         if (drainBatch) {
@@ -2176,8 +2181,30 @@ export class AgentManager {
         return entry;
     }
     /**
-     * Drop every queued message. Cancelling an agent is one "stop everything"
-     * verb: aborting the run also abandons the work you had lined up behind it.
+     * Keep the queue, but stop the cancelled turn from draining it.
+     *
+     * Cancel used to empty the queue outright, on the theory that "stop
+     * everything" includes the work lined up behind it. In practice the messages
+     * you queued are the ones you still want: stopping the run is how you make
+     * room for them, so wiping them destroys work the user typed. So stop now
+     * means exactly what it says — nothing new starts by itself — while the
+     * entries stay in the Queue track, ready to edit or send.
+     *
+     * Returns how many entries were held (0 when there was no queue), purely for
+     * logging.
+     */
+    holdSteerQueue(agentId) {
+        const agent = this.agents.get(agentId);
+        if (!agent) {
+            return 0;
+        }
+        agent.steerQueueHeld = true;
+        return agent.steerQueue.length;
+    }
+    /**
+     * Drop every queued message — the explicit "clear the queue" verb behind
+     * `agent.queue.clear`, and the way a closing agent sheds its queue. Cancel
+     * does NOT come through here; see `holdSteerQueue`.
      */
     clearSteerQueue(agentId) {
         const agent = this.agents.get(agentId);

package/dist/server/server/agent/lifecycle-command.d.ts

@@ -7,7 +7,7 @@ export interface LifecycleAgentManager {
     getAgent(agentId: string): LifecycleAgentSnapshot | null;
     hasInFlightRun(agentId: string): boolean;
     cancelAgentRun(agentId: string): Promise<boolean>;
-    clearSteerQueue(agentId: string): number;
+    holdSteerQueue(agentId: string): number;
     clearAgentAttention(agentId: string): Promise<void>;
     archiveAgent(agentId: string): Promise<{
         archivedAt: string;

package/dist/server/server/agent/lifecycle-command.js

@@ -5,12 +5,13 @@ export async function cancelAgentRunCommand(dependencies, agentId) {
         logger.trace({ agentId }, "cancelAgentRunCommand: agent not found");
         throw new Error(`Agent ${agentId} not found`);
     }
-    // Cancel is one "stop everything" verb: aborting the run also abandons the
-    // messages queued behind it. Runs before the not-running early return so an
-    // agent that somehow holds a queue while idle is still emptied.
-    const clearedQueue = agentManager.clearSteerQueue(agentId);
-    if (clearedQueue > 0) {
-        logger.debug({ agentId, clearedQueue }, "cancelAgentRunCommand: cleared queued messages");
+    // Cancel stops the run, not the plan: the queued messages stay, and the
+    // cancelled turn is stopped from draining them into a fresh turn. They sit in
+    // the Queue track ready to send. Runs BEFORE the interrupt so the hold is in
+    // place by the time the cancel finalizes the turn.
+    const heldQueue = agentManager.holdSteerQueue(agentId);
+    if (heldQueue > 0) {
+        logger.debug({ agentId, heldQueue }, "cancelAgentRunCommand: held queued messages");
     }
     const hasInFlightRun = agentManager.hasInFlightRun(agentId);
     if (!hasInFlightRun) {

package/dist/server/server/agent/providers/claude/agent.js

@@ -1543,21 +1543,35 @@ function isClaudeWorkflowToolName(name) {
 function isClaudeWorkflowTaskType(taskType) {
     return taskType === "local_workflow" || taskType === "workflow";
 }
-// Background-tasks track membership, decided by the SDK's task_type discriminant
-// ('shell' | 'subagent' | 'monitor' | 'workflow' | a raw string for unknown
-// types). Subagents and workflows have their own richer observed-subagent rows,
-// so the background track takes everything else the CLI backgrounds — shell
-// (Bash run_in_background), monitors, and any future/unknown background type.
-// Previously only the Bash tool name was recognized, so monitors and other
-// non-shell background tasks were silently dropped. Returns false for an absent
-// task_type (later task_progress omits it); those events fall back to the
-// remembered classification, so this only decides first-sighting (task_started /
+// An AI run that belongs in the observed-subagent track, by task_type. The CLI's
+// real discriminant values are local_bash / local_agent / remote_agent /
+// in_process_teammate / local_workflow / monitor_mcp / monitor_ws / mcp_task /
+// dream / auto_mode_scan — there is no "subagent". Otto excluded only that
+// never-emitted string, so every Agent/Task run (task_type "local_agent") also
+// matched isClaudeBackgroundTaskType and the background_tasks_changed level
+// signal — which the SDK documents as arriving BEFORE the task_started edge —
+// filed it as a background task too. The result was one sub-agent showing up
+// twice: once in each track, same title, seconds apart. "subagent" is kept as an
+// accepted alias so an older or renaming CLI still routes correctly.
+function isClaudeSubagentTaskType(taskType) {
+    return (taskType === "local_agent" ||
+        taskType === "remote_agent" ||
+        taskType === "in_process_teammate" ||
+        taskType === "subagent");
+}
+// Background-tasks track membership, decided by the SDK's task_type discriminant.
+// Sub-agents and workflows have their own richer observed-subagent rows, so the
+// background track takes everything else the CLI backgrounds — local_bash,
+// monitor_mcp/monitor_ws, mcp_task, dream, auto_mode_scan, and any future or
+// unknown background type. Returns false for an absent task_type (later
+// task_progress omits it); those events fall back to the remembered
+// classification, so this only decides first-sighting (task_started /
 // level-signal) routing.
 function isClaudeBackgroundTaskType(taskType) {
     if (!taskType) {
         return false;
     }
-    return taskType !== "subagent" && !isClaudeWorkflowTaskType(taskType);
+    return !isClaudeSubagentTaskType(taskType) && !isClaudeWorkflowTaskType(taskType);
 }
 // Frozen row label for an observed workflow run. workflow_name is meta.name from
 // the script (e.g. "spec"); prefix it so the observed track reads it as a
@@ -3908,7 +3922,7 @@ class ClaudeAgentSession {
         const isWorkflowStart = isClaudeWorkflowTaskType(message.task_type) || isClaudeWorkflowToolName(cachedTool?.name);
         const isSubagentStart = isClaudeSubagentToolName(cachedTool?.name) ||
             readObservedSubagentText(message.subagent_type) !== undefined ||
-            message.task_type === "subagent";
+            isClaudeSubagentTaskType(message.task_type);
         // Route to the background-tasks track anything the CLI backgrounds that is
         // not an observable AI run: an already-known background task, a Bash
         // run_in_background, a monitor, or any other non-subagent, non-workflow
@@ -3938,6 +3952,7 @@ class ClaudeAgentSession {
                 : message.subagent_type,
             description: message.description,
             status: "running",
+            classifiedAsSubagent: isWorkflowStart || isSubagentStart,
         });
     }
     /**
@@ -4080,6 +4095,27 @@ class ClaudeAgentSession {
         return (isClaudeBackgroundShellToolName(cachedTool?.name) ||
             this.backgroundShellKeyByTaskId.has(taskId));
     }
+    /**
+     * Whether a task_* message describes an observable AI run rather than a
+     * shell/monitor background task. `classifiedAsSubagent` is the caller's own
+     * verdict (task_started reads task_type); the rest re-derive membership for
+     * the later edges, which carry neither task_type nor subagent_type.
+     */
+    isObservableSubagentTask(input) {
+        if (input.classifiedAsSubagent === true) {
+            return true;
+        }
+        if (readObservedSubagentText(input.subAgentType) !== undefined) {
+            return true;
+        }
+        const cachedTool = input.toolUseId ? this.toolUseCache.get(input.toolUseId) : undefined;
+        // A Workflow orchestration run is observable too — recognize it directly by
+        // its cached tool so a task_progress/task_notification seen before (or
+        // without) task_started still routes to the observed row.
+        return (isClaudeSubagentToolName(cachedTool?.name) ||
+            isClaudeWorkflowToolName(cachedTool?.name) ||
+            this.observedKeyByTaskId.has(input.taskId));
+    }
     /**
      * Map a task_* system message onto the observed subagent's lifecycle. Only
      * subagent tasks qualify — shell/monitor/workflow background tasks are
@@ -4087,15 +4123,7 @@ class ClaudeAgentSession {
      */
     appendObservedSubagentTaskEvent(message, events, input) {
         void message;
-        const cachedTool = input.toolUseId ? this.toolUseCache.get(input.toolUseId) : undefined;
-        const isSubagentTask = readObservedSubagentText(input.subAgentType) !== undefined ||
-            isClaudeSubagentToolName(cachedTool?.name) ||
-            // A Workflow orchestration run is observable too — recognize it directly by
-            // its cached tool so a task_progress/task_notification seen before (or
-            // without) task_started still routes to the observed row.
-            isClaudeWorkflowToolName(cachedTool?.name) ||
-            this.observedKeyByTaskId.has(input.taskId);
-        if (!isSubagentTask) {
+        if (!this.isObservableSubagentTask(input)) {
             return;
         }
         const key = input.toolUseId ?? this.observedKeyByTaskId.get(input.taskId) ?? `task:${input.taskId}`;

package/dist/server/server/agent/providers/claude/claude-pricing.js

@@ -37,15 +37,19 @@ function rates(input, output) {
 // Exact model id (lowercased) → rates. Only the models we can price with
 // confidence; everything else is deliberately absent (→ undefined).
 const CLAUDE_MODEL_RATES = {
+    // Fable 5 — $10 in / $50 out, the tier above Opus.
+    "claude-fable-5": rates(10, 50),
     // Opus 5 and Opus 4.6+ — $5 in / $25 out. (The $15/$75 card belonged to
     // Opus 4.1 and earlier; every Opus we list here ships at the lower rate.)
     "claude-opus-5": rates(5, 25),
     "claude-opus-4-8": rates(5, 25),
     "claude-opus-4-7": rates(5, 25),
     "claude-opus-4-6": rates(5, 25),
+    "claude-opus-4-5": rates(5, 25),
     // Sonnet — $3 in / $15 out (standard, ≤200K context).
     "claude-sonnet-5": rates(3, 15),
     "claude-sonnet-4-6": rates(3, 15),
+    "claude-sonnet-4-5": rates(3, 15),
     // Haiku 4.5 — $1 in / $5 out. Both the plain and the dated API id.
     "claude-haiku-4-5": rates(1, 5),
     "claude-haiku-4-5-20251001": rates(1, 5),

package/dist/server/server/agent/providers/claude/model-manifest.d.ts

@@ -11,6 +11,29 @@ import type { AgentModelDefinition } from "../../agent-sdk-types.js";
  */
 export type ClaudeAutoModeSupport = "all" | "anthropic-api" | "none";
 export declare const CLAUDE_ULTRACODE_THINKING_OPTION_ID = "ultracode";
+/**
+ * The Claude models Otto offers, mirroring the Claude Code CLI's own catalog.
+ * Every field here is checked against that catalog (`context`, `capabilities`,
+ * `max_output_tokens`) rather than guessed from the model name:
+ *
+ * - A `[1m]` entry exists ONLY for models the CLI reports as `window:200000`
+ *   AND `supports_1m_beta`. Opus 4.7 / 4.8 / 5, Sonnet 5 and Fable 5 are
+ *   `native_1m` — the plain id already resolves to a 1M window, so a second
+ *   "1M" row would be a duplicate of the same model. Opus 4.5 is 200K with no
+ *   `supports_1m_beta`, so it gets no 1M row either. Ignore `supports_1m_suffix`
+ *   as a signal: nearly every model carries it, including Haiku 3.5.
+ * - `effortLevels` requires `effort` in the model's capabilities, and the
+ *   `xhigh` ladder additionally requires `xhigh_effort`. Opus 4.6 and Sonnet 4.6
+ *   carry `effort`/`max_effort` only; Opus 4.5, Sonnet 4.5 and Haiku 4.5 carry
+ *   neither, so they expose no thinking options at all.
+ * - `supportsFastMode` is Opus 4.8 and Opus 5 only. Opus 4.7's fast mode was
+ *   withdrawn (`speed: "fast"` now errors) and Opus 4.6 never had it.
+ *
+ * Deliberately NOT listed: retired ids (Sonnet 3.7, Haiku 3.5, Sonnet 3.5,
+ * Opus 3, Haiku 3); Opus 4.1, Opus 4 and Sonnet 4, which Anthropic has
+ * deprecated with retirement pending; and Mythos 5, which is invitation-only
+ * via Project Glasswing.
+ */
 export declare const CLAUDE_MODEL_MANIFEST: readonly [{
     readonly id: "claude-fable-5";
     readonly label: "Fable 5";
@@ -18,27 +41,11 @@ export declare const CLAUDE_MODEL_MANIFEST: readonly [{
     readonly contextWindowMaxTokens: 1000000;
     readonly effortLevels: readonly ["low", "medium", "high", "xhigh", "max"];
     readonly autoModeSupport: "all";
-}, {
-    readonly id: "claude-opus-5[1m]";
-    readonly label: "Opus 5 1M";
-    readonly description: "Opus 5 with 1M context window";
-    readonly contextWindowMaxTokens: 1000000;
-    readonly effortLevels: readonly ["low", "medium", "high", "xhigh", "max"];
-    readonly supportsFastMode: true;
-    readonly autoModeSupport: "all";
 }, {
     readonly id: "claude-opus-5";
     readonly label: "Opus 5";
     readonly description: "Opus 5 · Latest release";
     readonly isDefault: true;
-    readonly contextWindowMaxTokens: 200000;
-    readonly effortLevels: readonly ["low", "medium", "high", "xhigh", "max"];
-    readonly supportsFastMode: true;
-    readonly autoModeSupport: "all";
-}, {
-    readonly id: "claude-opus-4-8[1m]";
-    readonly label: "Opus 4.8 1M";
-    readonly description: "Opus 4.8 with 1M context window";
     readonly contextWindowMaxTokens: 1000000;
     readonly effortLevels: readonly ["low", "medium", "high", "xhigh", "max"];
     readonly supportsFastMode: true;
@@ -47,7 +54,7 @@ export declare const CLAUDE_MODEL_MANIFEST: readonly [{
     readonly id: "claude-opus-4-8";
     readonly label: "Opus 4.8";
     readonly description: "Opus 4.8 · Previous release";
-    readonly contextWindowMaxTokens: 200000;
+    readonly contextWindowMaxTokens: 1000000;
     readonly effortLevels: readonly ["low", "medium", "high", "xhigh", "max"];
     readonly supportsFastMode: true;
     readonly autoModeSupport: "all";
@@ -58,21 +65,12 @@ export declare const CLAUDE_MODEL_MANIFEST: readonly [{
     readonly contextWindowMaxTokens: 1000000;
     readonly effortLevels: readonly ["low", "medium", "high", "xhigh", "max"];
     readonly autoModeSupport: "all";
-}, {
-    readonly id: "claude-opus-4-7[1m]";
-    readonly label: "Opus 4.7 1M";
-    readonly description: "Opus 4.7 with 1M context window";
-    readonly contextWindowMaxTokens: 1000000;
-    readonly effortLevels: readonly ["low", "medium", "high", "xhigh", "max"];
-    readonly supportsFastMode: true;
-    readonly autoModeSupport: "all";
 }, {
     readonly id: "claude-opus-4-7";
     readonly label: "Opus 4.7";
     readonly description: "Opus 4.7 · Earlier release";
-    readonly contextWindowMaxTokens: 200000;
+    readonly contextWindowMaxTokens: 1000000;
     readonly effortLevels: readonly ["low", "medium", "high", "xhigh", "max"];
-    readonly supportsFastMode: true;
     readonly autoModeSupport: "all";
 }, {
     readonly id: "claude-opus-4-6[1m]";
@@ -80,15 +78,13 @@ export declare const CLAUDE_MODEL_MANIFEST: readonly [{
     readonly description: "Opus 4.6 with 1M context window";
     readonly contextWindowMaxTokens: 1000000;
     readonly effortLevels: readonly ["low", "medium", "high", "max"];
-    readonly supportsFastMode: true;
     readonly autoModeSupport: "anthropic-api";
 }, {
     readonly id: "claude-opus-4-6";
     readonly label: "Opus 4.6";
-    readonly description: "Opus 4.6 · Most capable for complex work";
+    readonly description: "Opus 4.6 · Legacy release";
     readonly contextWindowMaxTokens: 200000;
     readonly effortLevels: readonly ["low", "medium", "high", "max"];
-    readonly supportsFastMode: true;
     readonly autoModeSupport: "anthropic-api";
 }, {
     readonly id: "claude-sonnet-4-6[1m]";
@@ -104,6 +100,24 @@ export declare const CLAUDE_MODEL_MANIFEST: readonly [{
     readonly contextWindowMaxTokens: 200000;
     readonly effortLevels: readonly ["low", "medium", "high", "max"];
     readonly autoModeSupport: "anthropic-api";
+}, {
+    readonly id: "claude-opus-4-5";
+    readonly label: "Opus 4.5";
+    readonly description: "Opus 4.5 · Legacy release";
+    readonly contextWindowMaxTokens: 200000;
+    readonly autoModeSupport: "none";
+}, {
+    readonly id: "claude-sonnet-4-5[1m]";
+    readonly label: "Sonnet 4.5 1M";
+    readonly description: "Sonnet 4.5 with 1M context window";
+    readonly contextWindowMaxTokens: 1000000;
+    readonly autoModeSupport: "none";
+}, {
+    readonly id: "claude-sonnet-4-5";
+    readonly label: "Sonnet 4.5";
+    readonly description: "Sonnet 4.5 · Legacy release";
+    readonly contextWindowMaxTokens: 200000;
+    readonly autoModeSupport: "none";
 }, {
     readonly id: "claude-haiku-4-5";
     readonly label: "Haiku 4.5";

package/dist/server/server/agent/providers/claude/model-manifest.js

@@ -10,6 +10,29 @@ const CLAUDE_EFFORT_LABELS = {
     max: "Max",
 };
 export const CLAUDE_ULTRACODE_THINKING_OPTION_ID = "ultracode";
+/**
+ * The Claude models Otto offers, mirroring the Claude Code CLI's own catalog.
+ * Every field here is checked against that catalog (`context`, `capabilities`,
+ * `max_output_tokens`) rather than guessed from the model name:
+ *
+ * - A `[1m]` entry exists ONLY for models the CLI reports as `window:200000`
+ *   AND `supports_1m_beta`. Opus 4.7 / 4.8 / 5, Sonnet 5 and Fable 5 are
+ *   `native_1m` — the plain id already resolves to a 1M window, so a second
+ *   "1M" row would be a duplicate of the same model. Opus 4.5 is 200K with no
+ *   `supports_1m_beta`, so it gets no 1M row either. Ignore `supports_1m_suffix`
+ *   as a signal: nearly every model carries it, including Haiku 3.5.
+ * - `effortLevels` requires `effort` in the model's capabilities, and the
+ *   `xhigh` ladder additionally requires `xhigh_effort`. Opus 4.6 and Sonnet 4.6
+ *   carry `effort`/`max_effort` only; Opus 4.5, Sonnet 4.5 and Haiku 4.5 carry
+ *   neither, so they expose no thinking options at all.
+ * - `supportsFastMode` is Opus 4.8 and Opus 5 only. Opus 4.7's fast mode was
+ *   withdrawn (`speed: "fast"` now errors) and Opus 4.6 never had it.
+ *
+ * Deliberately NOT listed: retired ids (Sonnet 3.7, Haiku 3.5, Sonnet 3.5,
+ * Opus 3, Haiku 3); Opus 4.1, Opus 4 and Sonnet 4, which Anthropic has
+ * deprecated with retirement pending; and Mythos 5, which is invitation-only
+ * via Project Glasswing.
+ */
 export const CLAUDE_MODEL_MANIFEST = [
     {
         id: "claude-fable-5",
@@ -19,29 +42,11 @@ export const CLAUDE_MODEL_MANIFEST = [
         effortLevels: CLAUDE_EFFORT_LEVELS.xhigh,
         autoModeSupport: "all",
     },
-    {
-        id: "claude-opus-5[1m]",
-        label: "Opus 5 1M",
-        description: "Opus 5 with 1M context window",
-        contextWindowMaxTokens: 1000000,
-        effortLevels: CLAUDE_EFFORT_LEVELS.xhigh,
-        supportsFastMode: true,
-        autoModeSupport: "all",
-    },
     {
         id: "claude-opus-5",
         label: "Opus 5",
         description: "Opus 5 · Latest release",
         isDefault: true,
-        contextWindowMaxTokens: 200000,
-        effortLevels: CLAUDE_EFFORT_LEVELS.xhigh,
-        supportsFastMode: true,
-        autoModeSupport: "all",
-    },
-    {
-        id: "claude-opus-4-8[1m]",
-        label: "Opus 4.8 1M",
-        description: "Opus 4.8 with 1M context window",
         contextWindowMaxTokens: 1000000,
         effortLevels: CLAUDE_EFFORT_LEVELS.xhigh,
         supportsFastMode: true,
@@ -51,7 +56,7 @@ export const CLAUDE_MODEL_MANIFEST = [
         id: "claude-opus-4-8",
         label: "Opus 4.8",
         description: "Opus 4.8 · Previous release",
-        contextWindowMaxTokens: 200000,
+        contextWindowMaxTokens: 1000000,
         effortLevels: CLAUDE_EFFORT_LEVELS.xhigh,
         supportsFastMode: true,
         autoModeSupport: "all",
@@ -64,22 +69,12 @@ export const CLAUDE_MODEL_MANIFEST = [
         effortLevels: CLAUDE_EFFORT_LEVELS.xhigh,
         autoModeSupport: "all",
     },
-    {
-        id: "claude-opus-4-7[1m]",
-        label: "Opus 4.7 1M",
-        description: "Opus 4.7 with 1M context window",
-        contextWindowMaxTokens: 1000000,
-        effortLevels: CLAUDE_EFFORT_LEVELS.xhigh,
-        supportsFastMode: true,
-        autoModeSupport: "all",
-    },
     {
         id: "claude-opus-4-7",
         label: "Opus 4.7",
         description: "Opus 4.7 · Earlier release",
-        contextWindowMaxTokens: 200000,
+        contextWindowMaxTokens: 1000000,
         effortLevels: CLAUDE_EFFORT_LEVELS.xhigh,
-        supportsFastMode: true,
         autoModeSupport: "all",
     },
     {
@@ -88,16 +83,14 @@ export const CLAUDE_MODEL_MANIFEST = [
         description: "Opus 4.6 with 1M context window",
         contextWindowMaxTokens: 1000000,
         effortLevels: CLAUDE_EFFORT_LEVELS.standard,
-        supportsFastMode: true,
         autoModeSupport: "anthropic-api",
     },
     {
         id: "claude-opus-4-6",
         label: "Opus 4.6",
-        description: "Opus 4.6 · Most capable for complex work",
+        description: "Opus 4.6 · Legacy release",
         contextWindowMaxTokens: 200000,
         effortLevels: CLAUDE_EFFORT_LEVELS.standard,
-        supportsFastMode: true,
         autoModeSupport: "anthropic-api",
     },
     {
@@ -116,6 +109,27 @@ export const CLAUDE_MODEL_MANIFEST = [
         effortLevels: CLAUDE_EFFORT_LEVELS.standard,
         autoModeSupport: "anthropic-api",
     },
+    {
+        id: "claude-opus-4-5",
+        label: "Opus 4.5",
+        description: "Opus 4.5 · Legacy release",
+        contextWindowMaxTokens: 200000,
+        autoModeSupport: "none",
+    },
+    {
+        id: "claude-sonnet-4-5[1m]",
+        label: "Sonnet 4.5 1M",
+        description: "Sonnet 4.5 with 1M context window",
+        contextWindowMaxTokens: 1000000,
+        autoModeSupport: "none",
+    },
+    {
+        id: "claude-sonnet-4-5",
+        label: "Sonnet 4.5",
+        description: "Sonnet 4.5 · Legacy release",
+        contextWindowMaxTokens: 200000,
+        autoModeSupport: "none",
+    },
     {
         id: "claude-haiku-4-5",
         label: "Haiku 4.5",
@@ -249,7 +263,20 @@ function normalizeSingleSegmentClaudeModelId(familyValue, major, hasOneMillionCo
 function normalizeMajorMinorClaudeModelId(familyValue, major, minor, hasOneMillionContext) {
     const family = familyValue.toLowerCase();
     const suffix = hasOneMillionContext ? "[1m]" : "";
-    const candidate = `claude-${family}-${major}-${minor}${suffix}`;
-    return isClaudeManifestModelId(candidate) ? candidate : null;
+    // Fall back to the undecorated id when the manifest has no `[1m]` row, the
+    // same way normalizeSingleSegmentClaudeModelId does. Natively-1M models (Opus
+    // 4.7/4.8/5) ship one entry, but the CLI still accepts the decorated id and a
+    // persisted agent, personality, or team binding may carry it — without this
+    // fallback those resolve to null and silently lose their feature gates.
+    const candidates = [
+        `claude-${family}-${major}-${minor}${suffix}`,
+        `claude-${family}-${major}-${minor}`,
+    ];
+    for (const candidate of candidates) {
+        if (isClaudeManifestModelId(candidate)) {
+            return candidate;
+        }
+    }
+    return null;
 }
 //# sourceMappingURL=model-manifest.js.map

package/dist/server/server/speech/providers/local/sherpa/sherpa-runtime-env.d.ts

@@ -8,6 +8,26 @@ export declare function sherpaPlatformArch(platform?: NodeJS.Platform, arch?: st
 export declare function sherpaPlatformPackageName(platform?: NodeJS.Platform, arch?: string): string;
 export declare function sherpaLoaderEnvKey(platform?: NodeJS.Platform): SherpaLoaderEnvKey | null;
 export declare function prependEnvPath(existing: string | undefined, value: string): string;
+/**
+ * Redirect a path that landed inside `app.asar` to its `app.asar.unpacked`
+ * twin. In a packaged Electron build `require.resolve` returns a path inside
+ * the asar archive, which is a FILE, not a directory: nothing loads from it,
+ * and electron-builder unpacks native modules to `app.asar.unpacked` for
+ * exactly that reason.
+ *
+ * On Windows the un-redirected path is not merely useless, it is destructive.
+ * We prepend this directory to the daemon's own PATH so the native addon can
+ * find its DLLs, and every process the daemon spawns inherits it. Git's MSYS
+ * layer rewrites PATH from POSIX to Windows form for native children, and it
+ * gives up at the `app.asar` entry, silently dropping every entry after it,
+ * including the one holding node itself. The symptom is every git hook in an
+ * agent session dying in under a second with "'node' is not recognized", which
+ * reads as a failed check rather than as a truncated PATH. See
+ * docs/development.md.
+ *
+ * `exists` is injectable so the redirect is testable without a packaged build.
+ */
+export declare function resolveUnpackedLibDir(dir: string, exists?: (candidate: string) => boolean): string;
 export declare function resolveSherpaLoaderEnv(platform?: NodeJS.Platform, arch?: string): SherpaLoaderEnvResolution | null;
 export declare function applySherpaLoaderEnv(env: NodeJS.ProcessEnv, platform?: NodeJS.Platform, arch?: string): {
     changed: boolean;

package/dist/server/server/speech/providers/local/sherpa/sherpa-runtime-env.js

@@ -1,3 +1,4 @@
+import { existsSync } from "node:fs";
 import { createRequire } from "node:module";
 import path from "node:path";
 export function sherpaPlatformArch(platform = process.platform, arch = process.arch) {
@@ -26,6 +27,33 @@ export function prependEnvPath(existing, value) {
     }
     return [value, ...parts].join(path.delimiter);
 }
+/**
+ * Redirect a path that landed inside `app.asar` to its `app.asar.unpacked`
+ * twin. In a packaged Electron build `require.resolve` returns a path inside
+ * the asar archive, which is a FILE, not a directory: nothing loads from it,
+ * and electron-builder unpacks native modules to `app.asar.unpacked` for
+ * exactly that reason.
+ *
+ * On Windows the un-redirected path is not merely useless, it is destructive.
+ * We prepend this directory to the daemon's own PATH so the native addon can
+ * find its DLLs, and every process the daemon spawns inherits it. Git's MSYS
+ * layer rewrites PATH from POSIX to Windows form for native children, and it
+ * gives up at the `app.asar` entry, silently dropping every entry after it,
+ * including the one holding node itself. The symptom is every git hook in an
+ * agent session dying in under a second with "'node' is not recognized", which
+ * reads as a failed check rather than as a truncated PATH. See
+ * docs/development.md.
+ *
+ * `exists` is injectable so the redirect is testable without a packaged build.
+ */
+export function resolveUnpackedLibDir(dir, exists = existsSync) {
+    const match = dir.match(/^(.*app\.asar)([\\/].*)$/);
+    if (!match) {
+        return dir;
+    }
+    const unpacked = `${match[1]}.unpacked${match[2]}`;
+    return exists(unpacked) ? unpacked : dir;
+}
 export function resolveSherpaLoaderEnv(platform = process.platform, arch = process.arch) {
     const key = sherpaLoaderEnvKey(platform);
     if (!key) {
@@ -37,7 +65,7 @@ export function resolveSherpaLoaderEnv(platform = process.platform, arch = proce
         const pkgJson = require.resolve(`${packageName}/package.json`);
         return {
             key,
-            libDir: path.dirname(pkgJson),
+            libDir: resolveUnpackedLibDir(path.dirname(pkgJson)),
             packageName,
         };
     }

package/dist/server/web-ui/_expo/static/css/xterm-568e87a1b494979da9e20c7259569fdf.css

@@ -0,0 +1 @@
+.xterm{cursor:text;-webkit-user-select:none;user-select:none;position:relative}.xterm.focus,.xterm:focus{outline:none}.xterm .xterm-helpers{z-index:5;position:absolute;top:0}.xterm .xterm-helper-textarea{opacity:0;z-index:-5;white-space:nowrap;resize:none;border:0;width:0;height:0;margin:0;padding:0;position:absolute;top:0;left:-9999em;overflow:hidden}.xterm .composition-view{color:#fff;white-space:nowrap;z-index:1;background:#000;display:none;position:absolute}.xterm .composition-view.active{display:block}.xterm .xterm-viewport{cursor:default;position:absolute;inset:0;overflow-y:scroll}.xterm:not(.allow-transparency) .xterm-viewport{background-color:#000}.xterm .xterm-screen{position:relative}.xterm .xterm-screen canvas{position:absolute;top:0;left:0}.xterm.enable-mouse-events{cursor:default}.xterm.xterm-cursor-pointer,.xterm .xterm-cursor-pointer{cursor:pointer}.xterm.column-select.focus{cursor:crosshair}.xterm .xterm-accessibility:not(.debug),.xterm .xterm-message{z-index:10;color:#0000;pointer-events:none;position:absolute;inset:0}.xterm .xterm-accessibility-tree:not(.debug) ::selection{color:#0000}.xterm .xterm-accessibility-tree{-webkit-user-select:text;user-select:text;white-space:pre;font-family:monospace}.xterm .xterm-accessibility-tree>div{transform-origin:0;width:fit-content}.xterm .live-region{width:1px;height:1px;position:absolute;left:-9999px;overflow:hidden}.xterm-dim{opacity:1!important}.xterm-underline-1{text-decoration:underline}.xterm-underline-2{-webkit-text-decoration:underline double;text-decoration:underline double}.xterm-underline-3{-webkit-text-decoration:underline wavy;text-decoration:underline wavy}.xterm-underline-4{-webkit-text-decoration:underline dotted;text-decoration:underline dotted}.xterm-underline-5{-webkit-text-decoration:underline dashed;text-decoration:underline dashed}.xterm-overline{text-decoration:overline}.xterm-overline.xterm-underline-1{text-decoration:underline overline}.xterm-overline.xterm-underline-2{-webkit-text-decoration:overline double underline;text-decoration:overline double underline}.xterm-overline.xterm-underline-3{-webkit-text-decoration:overline wavy underline;text-decoration:overline wavy underline}.xterm-overline.xterm-underline-4{-webkit-text-decoration:overline dotted underline;text-decoration:overline dotted underline}.xterm-overline.xterm-underline-5{-webkit-text-decoration:overline dashed underline;text-decoration:overline dashed underline}.xterm-strikethrough{text-decoration:line-through}.xterm-screen .xterm-decoration-container .xterm-decoration{z-index:6;position:absolute}.xterm-screen .xterm-decoration-container .xterm-decoration.xterm-decoration-top-layer{z-index:7}.xterm-decoration-overview-ruler{z-index:8;pointer-events:none;position:absolute;top:0;right:0}.xterm-decoration-top{z-index:2;position:relative}.xterm .xterm-scrollable-element>.xterm-scrollbar{cursor:default}.xterm .xterm-scrollable-element>.xterm-scrollbar>.xterm-scra{cursor:pointer;background-color:var(--vscode-scrollbarSliderBackground,#64646466);-webkit-mask-image:url("data:image/svg+xml;utf8,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 11 11'><path d='M2.5 8.5 L5.5 2.5 L8.5 8.5 Z' fill='black'/></svg>");mask-image:url("data:image/svg+xml;utf8,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 11 11'><path d='M2.5 8.5 L5.5 2.5 L8.5 8.5 Z' fill='black'/></svg>");-webkit-mask-position:50%;mask-position:50%;-webkit-mask-size:100% 100%;mask-size:100% 100%;-webkit-mask-repeat:no-repeat;mask-repeat:no-repeat}.xterm .xterm-scrollable-element>.xterm-scrollbar>.xterm-scra.xterm-arrow-down{transform:rotate(180deg)}.xterm .xterm-scrollable-element>.xterm-visible{opacity:1;z-index:11;background:0 0;transition:opacity .1s linear}.xterm .xterm-scrollable-element>.xterm-invisible{opacity:0;pointer-events:none}.xterm .xterm-scrollable-element>.xterm-invisible.xterm-fade{transition:opacity .8s linear}.xterm .xterm-scrollable-element>.xterm-shadow{display:none;position:absolute}.xterm .xterm-scrollable-element>.xterm-shadow.xterm-shadow-top{width:100%;height:3px;box-shadow:var(--vscode-scrollbar-shadow,#000) 0 6px 6px -6px inset;display:block;top:0;left:3px}.xterm .xterm-scrollable-element>.xterm-shadow.xterm-shadow-left{width:3px;height:100%;box-shadow:var(--vscode-scrollbar-shadow,#000) 6px 0 6px -6px inset;display:block;top:3px;left:0}.xterm .xterm-scrollable-element>.xterm-shadow.xterm-shadow-top-left-corner{width:3px;height:3px;display:block;top:0;left:0}.xterm .xterm-scrollable-element>.xterm-shadow.xterm-shadow-top.xterm-shadow-left{box-shadow:var(--vscode-scrollbar-shadow,#000) 6px 0 6px -6px inset}