@vibedeckx/darwin-arm64 0.2.16 → 0.2.18

package/dist/bin.js

@@ -185613,6 +185613,7 @@ var mapAgentSession = (row) => ({
   permission_mode: row.permission_mode ?? void 0,
   agent_type: row.agent_type ?? void 0,
   title: row.title,
+  model: row.model ?? null,
   created_at: row.created_at,
   updated_at: row.updated_at,
   last_user_message_at: row.last_user_message_at,
@@ -185627,7 +185628,7 @@ var createAgentSessionRepos = (kdb, h) => ({
     // writes — this is what lets getLatestByBranch break ties
     // deterministically (see the schema.ts / sqlite.ts DDL comment on
     // agent_sessions).
-    create: async ({ id, project_id, branch, permission_mode, agent_type }) => {
+    create: async ({ id, project_id, branch, permission_mode, agent_type, model }) => {
       await kdb.insertInto("agent_sessions").values({
         id,
         project_id,
@@ -185635,6 +185636,7 @@ var createAgentSessionRepos = (kdb, h) => ({
         status: "running",
         permission_mode: permission_mode ?? "edit",
         agent_type: agent_type ?? "claude-code",
+        model: model ?? null,
         created_at: h.nowMs(),
         updated_at: h.nowMs()
       }).execute();
@@ -185677,6 +185679,9 @@ var createAgentSessionRepos = (kdb, h) => ({
     updateAgentType: async (id, agent_type) => {
       await kdb.updateTable("agent_sessions").set({ agent_type, updated_at: h.nowMs() }).where("id", "=", id).execute();
     },
+    updateModel: async (id, model) => {
+      await kdb.updateTable("agent_sessions").set({ model, updated_at: h.nowMs() }).where("id", "=", id).execute();
+    },
     updateTitle: async (id, title) => {
       await kdb.updateTable("agent_sessions").set({ title, updated_at: h.nowMs() }).where("id", "=", id).execute();
     },
@@ -186591,6 +186596,10 @@ var createDatabase = (dbPath) => {
       branch TEXT NOT NULL DEFAULT '',
       status TEXT NOT NULL DEFAULT 'running',
       title TEXT DEFAULT NULL,
+      -- Per-session agent model, e.g. 'opus' or 'gpt-5.6-sol'. NULL = use the
+      -- CLI's own default (no flag is passed). Never validated: an unknown
+      -- name is passed to the CLI and fails there.
+      model TEXT DEFAULT NULL,
       -- Millisecond-precision timestamps. CURRENT_TIMESTAMP is seconds-only,
       -- which lets two sessions tie on updated_at within the same second and
       -- corrupts the ordering used by getLatestByBranch. The 'YYYY-MM-DD
@@ -186863,6 +186872,10 @@ var createDatabase = (dbPath) => {
   if (!sessionInfoV6.some((col) => col.name === "favorited_at")) {
     db.exec("ALTER TABLE agent_sessions ADD COLUMN favorited_at INTEGER DEFAULT NULL");
   }
+  const sessionInfoV7 = db.prepare("PRAGMA table_info(agent_sessions)").all();
+  if (!sessionInfoV7.some((col) => col.name === "model")) {
+    db.exec("ALTER TABLE agent_sessions ADD COLUMN model TEXT DEFAULT NULL");
+  }
   db.exec(`
     CREATE INDEX IF NOT EXISTS idx_agent_sessions_project_branch
       ON agent_sessions(project_id, branch);
@@ -201310,8 +201323,11 @@ var UserInputParamsSchema = external_exports.looseObject({
 
 // src/protocol/codex/cli.ts
 var CROSS_REMOTE_MCP_TOKEN_ENV = "VIBEDECKX_CROSS_REMOTE_MCP_TOKEN";
-function buildCodexAppServerSpawnConfig(nativeBinary, crossRemoteMcp) {
+function buildCodexAppServerSpawnConfig(nativeBinary, crossRemoteMcp, model) {
   const args = ["app-server"];
+  if (model && model.trim()) {
+    args.push("-c", `model=${JSON.stringify(model.trim())}`);
+  }
   if (crossRemoteMcp) {
     args.push(
       "-c",
@@ -201396,11 +201412,16 @@ function withNpxFallback(nativeBinary, args) {
   }
   return { command: "npx", args: ["-y", CLAUDE_NPM_PACKAGE, ...args] };
 }
-function buildClaudeSessionSpawnConfig(nativeBinary, permissionMode, mcpConfigArg) {
+function buildClaudeSessionSpawnConfig(nativeBinary, permissionMode, mcpConfigArg, model) {
   const permissionFlag = permissionMode === "plan" ? "--permission-mode=plan" : "--dangerously-skip-permissions";
   const args = [
     ...STREAM_JSON_ARGS,
-    permissionFlag,
+    permissionFlag
+  ];
+  if (model && model.trim()) {
+    args.push("--model", model.trim());
+  }
+  args.push(
     // AskUserQuestion can't work over piped (non-TTY) stdin: claude resolves it
     // internally as "dismissed" before we can present a picker and wait for the
     // user. Disable it so the agent falls back to asking in plain text, which the
@@ -201408,7 +201429,7 @@ function buildClaudeSessionSpawnConfig(nativeBinary, permissionMode, mcpConfigAr
     "--disallowedTools",
     "AskUserQuestion",
     "--verbose"
-  ];
+  );
   if (mcpConfigArg) {
     args.push("--mcp-config", mcpConfigArg);
   }
@@ -202501,11 +202522,12 @@ var ClaudeCodeProvider = class {
   isAvailable() {
     return true;
   }
-  buildSpawnConfig(_cwd, permissionMode, crossRemoteMcp) {
+  buildSpawnConfig(_cwd, permissionMode, crossRemoteMcp, model) {
     return buildClaudeSessionSpawnConfig(
       this.detectBinary(),
       permissionMode,
-      crossRemoteMcp ? buildMcpConfigArg(crossRemoteMcp) : void 0
+      crossRemoteMcp ? buildMcpConfigArg(crossRemoteMcp) : void 0,
+      model
     );
   }
   parseStdoutLine(line, _sessionId) {
@@ -202695,9 +202717,9 @@ var CodexProvider = class _CodexProvider {
   isAvailable() {
     return true;
   }
-  buildSpawnConfig(_cwd, permissionMode, crossRemoteMcp) {
+  buildSpawnConfig(_cwd, permissionMode, crossRemoteMcp, model) {
     this.lastPermissionMode = permissionMode;
-    return buildCodexAppServerSpawnConfig(this.detectBinary(), crossRemoteMcp);
+    return buildCodexAppServerSpawnConfig(this.detectBinary(), crossRemoteMcp, model);
   }
   // ============ Task 5.5: parseStdoutLine — JSON-RPC message routing ============
   parseStdoutLine(line, sessionId) {
@@ -224595,6 +224617,21 @@ var TurnCompletionLedger = class {
 };
 
 // src/agent-session-manager.ts
+function buildStartupFailureMessage(agentType, stderrTail, stdoutTail) {
+  const provider = getProvider(agentType);
+  const name25 = provider.getDisplayName();
+  const details = [stdoutTail.trim(), stderrTail.trim()].filter(Boolean).join("\n");
+  const hint = stdoutTail.trim() ? void 0 : provider.getInstallHint?.();
+  let msg = `Couldn't start ${name25}.`;
+  if (hint) msg += `
+
+${hint}`;
+  if (details) msg += `
+
+Details:
+${details}`;
+  return msg;
+}
 var SUMMARY_TEXT_CAP = 1500;
 function extractLastAssistantText(entries) {
   for (let i = entries.length - 1; i >= 0; i--) {
@@ -224901,6 +224938,7 @@ var AgentSessionManager = class {
     await this.ensureResidentCapacity({ projectId, branch }, { force });
     const sessionId = opts.sessionId ?? randomUUID();
     const branchKey = branch ?? "";
+    const model = opts.model?.trim() ? opts.model.trim() : null;
     const absoluteWorktreePath = resolveWorktreePath(projectPath, branch);
     if (!skipDb) {
       await this.storage.agentSessions.create({
@@ -224908,7 +224946,8 @@ var AgentSessionManager = class {
         project_id: projectId,
         branch: branchKey,
         permission_mode: permissionMode,
-        agent_type: agentType
+        agent_type: agentType,
+        model
       });
     }
     if (!skipDb) {
@@ -224936,6 +224975,7 @@ var AgentSessionManager = class {
       permissionMode,
       crossRemoteMcp: opts.crossRemoteMcp,
       agentType,
+      model,
       completion: new TurnCompletionLedger(),
       graceTimer: null,
       eventChain: Promise.resolve(),
@@ -225026,7 +225066,7 @@ var AgentSessionManager = class {
       this.broadcastRaw(session.id, { finished: true });
       return;
     }
-    const config2 = provider.buildSpawnConfig(cwd, session.permissionMode, session.crossRemoteMcp);
+    const config2 = provider.buildSpawnConfig(cwd, session.permissionMode, session.crossRemoteMcp, session.model);
     if (config2.command !== "npx") {
       const agentVersion = getBinaryVersion(config2.command);
       console.log(`[AgentSession] ${provider.getDisplayName()} version: ${agentVersion ?? "unknown (--version probe failed)"}`);
@@ -225034,6 +225074,7 @@ var AgentSessionManager = class {
       console.log(`[AgentSession] ${provider.getDisplayName()} running via npx (version resolved at spawn by npm)`);
     }
     session.producedOutput = false;
+    session.unparsedStdoutTail = "";
     this.resetCompletion(session);
     let stderrTail = "";
     let spawnFailed = false;
@@ -225082,7 +225123,7 @@ var AgentSessionManager = class {
           try {
             await this.pushEntry(session.id, {
               type: "error",
-              message: this.buildStartupFailureMessage(session.agentType, stderrTail),
+              message: buildStartupFailureMessage(session.agentType, stderrTail, session.unparsedStdoutTail ?? ""),
               timestamp: Date.now()
             }, true);
           } catch (err) {
@@ -225115,7 +225156,7 @@ var AgentSessionManager = class {
       const isNotFound = error48.code === "ENOENT";
       this.pushEntry(session.id, {
         type: "error",
-        message: isNotFound ? this.buildStartupFailureMessage(session.agentType, stderrTail) : error48.message,
+        message: isNotFound ? buildStartupFailureMessage(session.agentType, stderrTail, session.unparsedStdoutTail ?? "") : error48.message,
         timestamp: Date.now()
       }, true).catch((err) => {
         console.error(`[AgentSession] Failed to push spawn-error entry for ${session.id}:`, err);
@@ -225240,6 +225281,8 @@ var AgentSessionManager = class {
       const events = provider.parseStdoutLine(line, session.id);
       if (events.length > 0) {
         session.producedOutput = true;
+      } else {
+        session.unparsedStdoutTail = ((session.unparsedStdoutTail ?? "") + line + "\n").slice(-4e3);
       }
       for (const event of events) {
         await this.processAgentEvent(session.id, event);
@@ -225493,25 +225536,6 @@ var AgentSessionManager = class {
   /**
    * Push a new entry with ADD patch
    */
-  /**
-   * Build a user-facing message for when an agent process fails to start.
-   * Includes the provider's install hint plus any captured stderr tail.
-   */
-  buildStartupFailureMessage(agentType, stderrTail) {
-    const provider = getProvider(agentType);
-    const name25 = provider.getDisplayName();
-    const hint = provider.getInstallHint?.();
-    let msg = `Couldn't start ${name25}.`;
-    if (hint) msg += `
-
-${hint}`;
-    const details = stderrTail.trim();
-    if (details) msg += `
-
-Details:
-${details}`;
-    return msg;
-  }
   /**
    * Allocate the next entry index, record the message in the in-memory store,
    * and build its ADD replay patch. Shared by `pushEntry` and `pushTurnEnd` so
@@ -226003,12 +226027,19 @@ ${details}`;
     session.store.currentAssistantIndex = null;
     session.buffer = "";
     this.resetCompletion(session);
+    const previousAgentType = session.agentType;
     getProvider(session.agentType).onSessionDestroyed?.(sessionId);
     session.agentType = agentType;
     if (!session.skipDb) await this.storage.agentSessions.updateAgentType(sessionId, agentType);
+    const clearedModel = session.model;
+    if (clearedModel !== null) {
+      session.model = null;
+      if (!session.skipDb) await this.storage.agentSessions.updateModel(sessionId, null);
+    }
+    const agentDisplayName = (t) => t === "codex" ? "Codex" : "Claude Code";
     await this.pushEntry(sessionId, {
       type: "system",
-      content: `Coding agent switched to ${agentType === "codex" ? "Codex" : "Claude Code"}.`,
+      content: `Coding agent switched to ${agentDisplayName(agentType)}.` + (clearedModel !== null ? ` Model reset to the default (\`${clearedModel}\` was set for ${agentDisplayName(previousAgentType)}).` : ""),
       timestamp: Date.now()
     });
     session.dormant = true;
@@ -226320,6 +226351,7 @@ ${details}`;
         skipDb: false,
         permissionMode,
         agentType: dbSession.agent_type || "claude-code",
+        model: dbSession.model ?? null,
         completion: new TurnCompletionLedger(),
         graceTimer: null,
         eventChain: Promise.resolve(),
@@ -226381,13 +226413,16 @@ ${details}`;
     const branch = source?.branch ?? (sourceRow.branch || null);
     const permissionMode = source?.permissionMode ?? (sourceRow?.permission_mode === "plan" ? "plan" : "edit");
     const agentType = agentTypeOverride ?? source?.agentType ?? (sourceRow?.agent_type || "claude-code");
+    const sourceAgentType = source?.agentType ?? (sourceRow?.agent_type || "claude-code");
+    const model = agentType === sourceAgentType ? source?.model ?? sourceRow?.model ?? null : null;
     const newId = opts.sessionId ?? randomUUID();
     await this.storage.agentSessions.create({
       id: newId,
       project_id: projectId,
       branch: branch ?? "",
       permission_mode: permissionMode,
-      agent_type: agentType
+      agent_type: agentType,
+      model
     });
     await this.storage.agentSessions.updateStatusPreservingTimestamp(newId, "stopped");
     for (const row of entryRows) {
@@ -226422,6 +226457,7 @@ ${details}`;
       skipDb: false,
       permissionMode,
       agentType,
+      model,
       completion: new TurnCompletionLedger(),
       graceTimer: null,
       eventChain: Promise.resolve(),
@@ -226865,6 +226901,8 @@ var WATCH_WINDOW_MS = 30 * 60 * 1e3;
 var WATCH_EXTEND_THROTTLE_MS = 5 * 60 * 1e3;
 var MAX_TRACKED_SESSIONS = 5e3;
 var SYNC_INTERVAL_MS = 6e4;
+var COMPLETION_NUDGE_DEBOUNCE_MS = 250;
+var MAX_CONCURRENT_SERVERS = 4;
 var MAX_PAGES_PER_SYNC = 50;
 function localWorkflowRunId(remoteServerId, projectId, remoteRunId) {
   return `remote-${remoteServerId}-${projectId}-${remoteRunId}`;
@@ -226877,10 +226915,19 @@ var RemoteNotificationSync = class {
   notificationService;
   reverseConnectManager;
   proxy;
+  nudgeDebounceMs;
   timer = null;
   stopped = false;
-  /** Serializes sweeps so a came-online sweep can't interleave with the tick. */
-  chain = Promise.resolve();
+  /**
+   * Per-server work queues. Two syncs of the SAME server must not interleave
+   * (they share its cursors), but two different servers have nothing in
+   * common — serializing them globally only meant an unreachable worker's
+   * retry budget delayed every healthy worker's import behind it.
+   */
+  serverChains = /* @__PURE__ */ new Map();
+  /** Local session ids whose completion is waiting for the debounce to fire. */
+  nudgeSessions = /* @__PURE__ */ new Set();
+  nudgeTimer = null;
   /**
    * Remote sessions this front currently believes are RUNNING. They are polled
    * regardless of their persisted watch window: an agent turn can easily run
@@ -226896,6 +226943,7 @@ var RemoteNotificationSync = class {
     this.notificationService = deps.notificationService;
     this.reverseConnectManager = deps.reverseConnectManager;
     this.proxy = deps.proxy ?? proxyToRemoteAuto;
+    this.nudgeDebounceMs = deps.nudgeDebounceMs ?? COMPLETION_NUDGE_DEBOUNCE_MS;
   }
   start() {
     if (this.timer || this.stopped) return;
@@ -226908,6 +226956,9 @@ var RemoteNotificationSync = class {
     this.stopped = true;
     if (this.timer) clearInterval(this.timer);
     this.timer = null;
+    if (this.nudgeTimer) clearTimeout(this.nudgeTimer);
+    this.nudgeTimer = null;
+    this.nudgeSessions.clear();
     this.activeRemoteSessions.clear();
     this.lastWatchExtend.clear();
   }
@@ -226931,6 +226982,7 @@ var RemoteNotificationSync = class {
     } else if (event.type === "session:taskCompleted" || event.type === "session:finished") {
       this.activeRemoteSessions.delete(sessionId);
     }
+    if (isTurnTerminalEvent(event)) this.scheduleCompletionSync(sessionId);
     const now2 = Date.now();
     const last = this.lastWatchExtend.get(sessionId) ?? 0;
     if (now2 - last < WATCH_EXTEND_THROTTLE_MS) return;
@@ -226944,10 +226996,65 @@ var RemoteNotificationSync = class {
       if (now2 - at > WATCH_WINDOW_MS) this.lastWatchExtend.delete(sessionId);
     }
   }
-  /** Fire-and-forget sweep, queued behind any in-flight one. */
+  /**
+   * Fire-and-forget sweep. No longer a global queue: the per-server chains
+   * inside syncAll/syncServer own the mutual exclusion that actually matters
+   * (one server's cursors), so callers here don't wait on unrelated servers.
+   */
   enqueue(work) {
     if (this.stopped) return;
-    this.chain = this.chain.then(work).catch((err) => console.warn("[RemoteNotifications] sync failed:", err));
+    void work().catch((err) => console.warn("[RemoteNotifications] sync failed:", err));
+  }
+  /**
+   * Queue `work` behind whatever is already running for this server, and
+   * nothing else. Errors are contained so one server's failure can't reject
+   * an unrelated caller awaiting the same chain.
+   */
+  runOnServer(remoteServerId, work) {
+    const previous = this.serverChains.get(remoteServerId) ?? Promise.resolve();
+    const next = previous.then(work).catch((err) => console.warn(`[RemoteNotifications] sync for ${remoteServerId} failed:`, err)).finally(() => {
+      if (this.serverChains.get(remoteServerId) === next) this.serverChains.delete(remoteServerId);
+    });
+    this.serverChains.set(remoteServerId, next);
+    return next;
+  }
+  /**
+   * Debounced, per-server pull for sessions whose turn just ended.
+   *
+   * Coalescing state is intentionally separate from `lastWatchExtend`: that
+   * map throttles DB writes on a 5-minute scale, which would swallow nearly
+   * every completion.
+   */
+  scheduleCompletionSync(localSessionId) {
+    if (this.stopped) return;
+    this.nudgeSessions.add(localSessionId);
+    if (this.nudgeTimer) return;
+    this.nudgeTimer = setTimeout(() => {
+      this.nudgeTimer = null;
+      const batch = [...this.nudgeSessions];
+      this.nudgeSessions.clear();
+      void this.syncSessions(batch).catch(
+        (err) => console.warn("[RemoteNotifications] completion sync failed:", err)
+      );
+    }, this.nudgeDebounceMs);
+    this.nudgeTimer.unref?.();
+  }
+  /**
+   * Pull exactly the mappings named, grouped per server. Deliberately does NOT
+   * go through `candidates()`: an explicit completion signal outranks the
+   * watch window, which may well have lapsed during a long turn.
+   */
+  async syncSessions(localSessionIds) {
+    const byServer = /* @__PURE__ */ new Map();
+    for (const localSessionId of localSessionIds) {
+      const mapping = await this.storage.remoteSessionMappings.getByLocal(localSessionId).catch(() => void 0);
+      if (!mapping) continue;
+      const group2 = byServer.get(mapping.remote_server_id);
+      if (group2) group2.push(mapping);
+      else byServer.set(mapping.remote_server_id, [mapping]);
+    }
+    if (byServer.size === 0) return;
+    await this.dispatchByServer([...byServer]);
   }
   /** Extend a mapping's periodic-poll window (create / send / workflow start / live activity). */
   async extendWatch(localSessionId) {
@@ -226997,17 +227104,33 @@ var RemoteNotificationSync = class {
     );
     return true;
   }
-  /** Sweep every mapped remote server. */
+  /** Sweep every mapped remote server, up to MAX_CONCURRENT_SERVERS at a time. */
   async syncAll(opts) {
     const mappings = await this.candidates(opts);
-    for (const [serverId, group2] of groupByServer(mappings)) {
-      await this.syncMappings(serverId, group2);
-    }
+    await this.dispatchByServer([...groupByServer(mappings)]);
   }
   /** Sweep one server — used when a reverse connection comes online. */
   async syncServer(remoteServerId, opts) {
     const mappings = (await this.candidates(opts)).filter((m2) => m2.remote_server_id === remoteServerId);
-    if (mappings.length > 0) await this.syncMappings(remoteServerId, mappings);
+    if (mappings.length > 0) await this.dispatchByServer([[remoteServerId, mappings]]);
+  }
+  /**
+   * Run one group per server, bounded. Each group still queues on its own
+   * server chain, so this composes with a nudge that arrives mid-sweep: same
+   * server → it waits its turn; different server → it runs immediately.
+   */
+  async dispatchByServer(groups) {
+    let next = 0;
+    const workers = Array.from(
+      { length: Math.min(MAX_CONCURRENT_SERVERS, groups.length) },
+      async () => {
+        for (let i = next++; i < groups.length; i = next++) {
+          const [serverId, group2] = groups[i];
+          await this.runOnServer(serverId, () => this.syncMappings(serverId, group2));
+        }
+      }
+    );
+    await Promise.all(workers);
   }
   /**
    * Mappings this sweep should poll: the persisted watch set, plus any session
@@ -227154,6 +227277,10 @@ var RemoteNotificationSync = class {
     };
   }
 };
+function isTurnTerminalEvent(event) {
+  if (event.type === "session:taskCompleted" || event.type === "session:finished") return true;
+  return event.type === "session:status" && (event.status === "stopped" || event.status === "error");
+}
 function groupByServer(mappings) {
   const groups = /* @__PURE__ */ new Map();
   for (const mapping of mappings) {
@@ -227166,7 +227293,7 @@ function groupByServer(mappings) {
 
 // src/remote-agent-sessions.ts
 async function createRemoteAgentSession(deps, params) {
-  const { projectId, agentMode, remoteConfig, branch, permissionMode, agentType, force, userId } = params;
+  const { projectId, agentMode, remoteConfig, branch, permissionMode, agentType, model, force, userId } = params;
   const remoteSessionId = randomUUID3();
   const localSessionId = `remote-${agentMode}-${projectId}-${remoteSessionId}`;
   const crossRemoteMcp = await mintCrossRemoteMcpConfig(
@@ -227188,7 +227315,7 @@ async function createRemoteAgentSession(deps, params) {
       remoteConfig.server_api_key || "",
       "POST",
       `/api/path/agent-sessions/new`,
-      { path: remoteConfig.remote_path, branch, permissionMode, agentType, force, sessionId: remoteSessionId, crossRemoteMcp },
+      { path: remoteConfig.remote_path, branch, permissionMode, agentType, force, sessionId: remoteSessionId, crossRemoteMcp, model },
       { reverseConnectManager: deps.reverseConnectManager ?? void 0 }
     );
     if (!result.ok) {
@@ -229914,12 +230041,18 @@ function selfReportSection(report) {
   if (!report) return null;
   return [
     "\n## Author's self-report (unverified)",
-    "The implementing agent described its own work as follows. Treat every claim as unverified \u2014 check each one against the actual code, and look for problems the self-report does not mention.",
+    "The implementing agent described its own work as follows. Treat every claim as unverified \u2014 verify first the claims that bear on whether the work achieves its goal, and look for problems the self-report does not mention.",
     "<author-self-report>",
     report,
     "</author-self-report>"
   ].join("\n");
 }
+var VERDICT_INSTRUCTIONS = [
+  "\nEnd your final message with:",
+  "1. Verdict \u2014 exactly one of: ship / needs-changes / cannot-verify. Use cannot-verify when you could not gather enough evidence to judge, rather than guessing.",
+  "2. Blocking findings \u2014 what must change before shipping, each specific and actionable (say explicitly when there are none).",
+  "3. Non-blocking notes \u2014 style and polish, briefly, clearly separated from the blocking list."
+];
 function buildReviewerPrompt(opts) {
   const intent = opts.originalIntent !== opts.taskContext ? opts.originalIntent : null;
   const brief = opts.intentBrief || null;
@@ -229958,7 +230091,12 @@ Confine your review to these files and changes. Other uncommitted or pre-existin
     "- Inspect the actual workspace state yourself: read the relevant files, run `git diff`, `git status` and `git log`.",
     reviewTargetPromptLine(opts.target),
     noDiffWithAnalysis ? "- Judge correctness and completeness against the task. For this analysis/plan turn the work under review is the reasoning and the proposal, not code quality of a diff. Be specific: reference files and lines." : "- Judge correctness, completeness against the task, and code quality. Be specific: reference files and lines.",
-    "\nEnd your final message with a clear, actionable list of feedback items \u2014 or state explicitly that the work looks good.",
+    // These two only make sense against a distilled brief: tier 2 has no
+    // [settled]/[tentative] marks and no stated scope, and implying it does
+    // would suppress findings on the strength of data that doesn't exist.
+    brief ? "- Where the brief marks a decision, non-goal, or accepted limitation as [settled], do not re-raise the choice itself as a finding; DO report concrete consequences it causes \u2014 failure of the core goal, or a correctness, security, or data loss problem. Items marked [tentative] (or unmarked) get normal review. A violated hard constraint is always blocking." : null,
+    brief ? "- Do not propose enhancements beyond the brief's stated scope \u2014 scope expansion is a product decision, not a review finding." : null,
+    ...VERDICT_INSTRUCTIONS,
     brief ? opts.authorSelfReport ? "\n(review context: distilled intent brief + author self-report + live workspace)" : "\n(review context: distilled intent brief + live workspace)" : hasExcerpt ? "\n(review context: deterministic excerpt of the source conversation + live workspace)" : "\n(review context: live workspace only \u2014 the source conversation was unavailable)"
   ].filter((l) => l !== null).join("\n");
 }
@@ -229982,7 +230120,7 @@ ${opts.reviewFocus}` : null,
     "- Check for regressions and remaining correctness or test gaps.",
     "- Do NOT modify files \u2014 remain in read-only review mode.",
     reviewTargetPromptLine(opts.target),
-    "- End with actionable feedback, or explicitly state that it looks good."
+    ...VERDICT_INSTRUCTIONS
   ].filter((line) => line !== null).join("\n");
 }
 function buildFeedbackMessage(feedback) {
@@ -235587,6 +235725,14 @@ function resolveUserId(authResult) {
 
 // src/routes/agent-session-routes.ts
 import { randomUUID as randomUUID15 } from "crypto";
+
+// src/protocol/model-suggestions.ts
+var MODEL_SUGGESTIONS = {
+  "claude-code": ["opus", "sonnet", "haiku", "fable"],
+  codex: ["gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna", "gpt-5.5"]
+};
+
+// src/routes/agent-session-routes.ts
 async function resolveProjectPath(projectId, storage) {
   if (projectId.startsWith("path:")) {
     return projectId.slice(5);
@@ -235650,6 +235796,7 @@ var routes12 = async (fastify2) => {
           status: session?.status || "stopped",
           permissionMode: session?.permissionMode || "edit",
           agentType: session?.agentType || "claude-code",
+          model: session?.model ?? null,
           title: dbRow?.title ?? null
         },
         messages
@@ -235660,15 +235807,23 @@ var routes12 = async (fastify2) => {
     fastify2.agentSessionManager.emitSessionTitle(projectId, branch, sessionId, title);
   }
   fastify2.get("/api/agent-providers", async (_req, reply) => {
-    const providers2 = getAllProviders().map((provider) => ({
-      type: provider.getAgentType(),
-      displayName: provider.getDisplayName(),
-      available: provider.isAvailable?.() ?? provider.detectBinary() !== null
-    }));
+    const providers2 = getAllProviders().map((provider) => {
+      const type = provider.getAgentType();
+      return {
+        type,
+        displayName: provider.getDisplayName(),
+        available: provider.isAvailable?.() ?? provider.detectBinary() !== null,
+        // Suggestions only. The picker also accepts free text, and nothing
+        // validates against this list — the session's real capabilities depend
+        // on the CLI version and account tier of whichever machine spawns it,
+        // which this server cannot know.
+        models: MODEL_SUGGESTIONS[type] ?? []
+      };
+    });
     return reply.code(200).send({ providers: providers2 });
   });
   fastify2.post("/api/path/agent-sessions", async (req, reply) => {
-    const { path: projectPath, branch, permissionMode, agentType, force } = req.body;
+    const { path: projectPath, branch, permissionMode, agentType, force, model } = req.body;
     if (!projectPath) {
       return reply.code(400).send({ error: "Path is required" });
     }
@@ -235716,6 +235871,7 @@ var routes12 = async (fastify2) => {
           status: effectiveStatus,
           permissionMode: session?.permissionMode || "edit",
           agentType: session?.agentType || "claude-code",
+          model: session?.model ?? null,
           processAlive: session ? fastify2.agentSessionManager.getSessionProcessAlive(sessionId) : false
         },
         messages
@@ -235760,7 +235916,7 @@ var routes12 = async (fastify2) => {
     }
   );
   fastify2.post("/api/path/agent-sessions/new", async (req, reply) => {
-    const { path: projectPath, branch, permissionMode, agentType, force, sessionId, crossRemoteMcp } = req.body;
+    const { path: projectPath, branch, permissionMode, agentType, force, sessionId, crossRemoteMcp, model } = req.body;
     if (!projectPath) {
       return reply.code(400).send({ error: "Path is required" });
     }
@@ -235788,7 +235944,7 @@ var routes12 = async (fastify2) => {
         agentType || "claude-code",
         false,
         force === true,
-        { sessionId, crossRemoteMcp }
+        { sessionId, crossRemoteMcp, model }
       );
       const session = fastify2.agentSessionManager.getSession(createdSessionId);
       return reply.code(200).send({
@@ -235799,6 +235955,7 @@ var routes12 = async (fastify2) => {
           status: session?.status || "running",
           permissionMode: session?.permissionMode || "edit",
           agentType: session?.agentType || "claude-code",
+          model: session?.model ?? null,
           processAlive: session ? fastify2.agentSessionManager.getSessionProcessAlive(session.id) : false
         },
         messages: []
@@ -235901,7 +236058,7 @@ var routes12 = async (fastify2) => {
     if (!project) {
       return reply.code(404).send({ error: "Project not found" });
     }
-    const { branch, permissionMode, agentType } = req.body;
+    const { branch, permissionMode, agentType, model } = req.body;
     let agentMode = project.agent_mode;
     let useRemoteAgent = agentMode !== "local";
     if (useRemoteAgent && agentMode === "remote") {
@@ -235936,7 +236093,7 @@ var routes12 = async (fastify2) => {
           remoteConfig.server_api_key || "",
           "POST",
           `/api/path/agent-sessions`,
-          { path: remoteConfig.remote_path, branch, permissionMode, agentType }
+          { path: remoteConfig.remote_path, branch, permissionMode, agentType, model }
         );
         console.log(`[API] Remote proxy result: ok=${result.ok}, status=${result.status}, data=${JSON.stringify(result.data).substring(0, 500)}`);
         if (result.ok) {
@@ -236014,6 +236171,7 @@ var routes12 = async (fastify2) => {
           status: effectiveStatus,
           permissionMode: session?.permissionMode || "edit",
           agentType: session?.agentType || "claude-code",
+          model: session?.model ?? null,
           processAlive: session ? fastify2.agentSessionManager.getSessionProcessAlive(sessionId) : false
         },
         messages
@@ -236030,7 +236188,7 @@ var routes12 = async (fastify2) => {
     if (!project) {
       return reply.code(404).send({ error: "Project not found" });
     }
-    const { branch, permissionMode, agentType, force } = req.body;
+    const { branch, permissionMode, agentType, force, model } = req.body;
     const agentMode = project.agent_mode;
     const useRemoteAgent = agentMode !== "local";
     if (useRemoteAgent) {
@@ -236056,7 +236214,8 @@ var routes12 = async (fastify2) => {
             permissionMode: permissionMode || "edit",
             agentType,
             force: force === true,
-            userId
+            userId,
+            model
           }
         );
         if (created.ok) {
@@ -236089,7 +236248,7 @@ var routes12 = async (fastify2) => {
         agentType || "claude-code",
         false,
         force === true,
-        { sessionId: preSessionId, crossRemoteMcp }
+        { sessionId: preSessionId, crossRemoteMcp, model }
       );
       const session = fastify2.agentSessionManager.getSession(sessionId);
       return reply.code(200).send({
@@ -236100,6 +236259,7 @@ var routes12 = async (fastify2) => {
           status: session?.status || "running",
           permissionMode: session?.permissionMode || "edit",
           agentType: session?.agentType || "claude-code",
+          model: session?.model ?? null,
           processAlive: session ? fastify2.agentSessionManager.getSessionProcessAlive(sessionId) : false
         },
         messages: []
@@ -236156,6 +236316,7 @@ var routes12 = async (fastify2) => {
           status: session.status,
           permissionMode: session.permissionMode,
           agentType: session.agentType || "claude-code",
+          model: session.model ?? null,
           processAlive: fastify2.agentSessionManager.getSessionProcessAlive(session.id)
         },
         messages
@@ -237518,77 +237679,295 @@ var command_routes_default = (0, import_fastify_plugin19.default)(routes18, { na
 var import_fastify_plugin20 = __toESM(require_plugin2(), 1);
 
 // src/utils/review-brief.ts
-var AI_TIMEOUT_MS2 = 15e3;
-var PER_MESSAGE_MAX_CHARS = 1500;
-var TOTAL_INPUT_MAX_CHARS = 24e3;
-var HEAD_CHARS = 8e3;
-var TAIL_CHARS = 15e3;
+var BASE_TIMEOUT_MS = 15e3;
+var JUDGMENT_TIMEOUT_MS = 9e4;
+var TIMEOUT_MS_PER_1K_CHARS = 120;
+var MAX_TIMEOUT_MS = 6e4;
+var PER_MESSAGE_MAX_CHARS = 6e3;
+var COMPACT_TRIGGER_CHARS = 12e4;
+var COMPACT_SLICE_BUDGETS = [96e3, 24e3, 6e3];
+var RECENT_VERBATIM_RATIO = 0.3;
+var MIN_RECENT_VERBATIM_CHARS = 2e3;
+var MAX_FOLD_CALLS = 8;
+var COMPACT_SUMMARY_MAX_CHARS = 6e3;
+var COMPACT_SUMMARY_MAX_TOKENS = 3e3;
+var REVERSALS_MAX_CHARS = 2e3;
+var REVERSALS_MAX_TOKENS = 2e3;
 var BRIEF_MAX_CHARS = 4e3;
+var BRIEF_MAX_TOKENS = 3e3;
+var REVERSAL_MIN_CHARS = 6e3;
+var SEP = "\n\n";
+var COMPACTION_NOTE = "_Note: the source conversation was compressed before distillation \u2014 older turns were summarized rather than read verbatim._";
+var BRIEF_TRUNCATION_NOTE = "\n\n_[brief truncated at the length limit \u2014 treat it as incomplete]_";
 var SYSTEM_PROMPT2 = [
   "You distill a coding-agent conversation into an intent brief for an independent code reviewer.",
   "The reviewer will NOT see the conversation \u2014 only your brief plus the actual code, so capture what the code alone cannot show:",
+  "0. Dominant question \u2014 1 or 2 sentences: the single question the reviewer must answer to judge this work (what must actually work, toward what goal). Derive it from the conversation's own emphasis \u2014 what the user repeated, finally decided, or reversed course over \u2014 never from your own engineering judgment. If the conversation does not clearly support one, write 'unclear' instead of inventing one.",
   "1. The original request and its goal.",
-  "2. Constraints and explicit user decisions, including approaches the user rejected.",
-  "3. Trade-offs or limitations that were acknowledged and accepted.",
-  "Do NOT include the agent's reasoning, self-assessment, or claims that the work is correct or complete \u2014 the reviewer must judge that independently.",
-  "Write concise markdown bullets under those numbered headings, under 400 words total, in the same language as the conversation. Reply with the brief only."
+  "2. Hard constraints \u2014 requirements whose violation the user would treat as failure regardless of everything else (security, data loss, an explicit compatibility or behaviour promise). Only list ones the user actually stated; an empty section is fine.",
+  "3. Constraints and decisions, including approaches that were rejected and why. A rejection counts whether the user made it or the agent made it under a principle the user set.",
+  "4. Non-goals, trade-offs and limitations that were acknowledged and accepted.",
+  "5. Behaviour established by actually running something \u2014 observed output, exit codes, error text, which stream an error arrived on. Report what was observed; that is evidence the reviewer cannot recover by reading code.",
+  "Tag every item under headings 3 and 4 with exactly [settled] or [tentative] \u2014 the bare tag; keep any justification to a few words or omit it, always after the tag, never inside the brackets. Settled ONLY when the user explicitly confirmed it, instructed it, or reversed course to it; an agent proposal the user never answered is tentative \u2014 silence is never settled.",
+  "Within each heading, order items by how much they matter to the dominant question, not by when they were said.",
+  "Later statements supersede earlier ones on the same topic. An option explored at length early is often rejected later in a single sentence, and the rejection wins. Where the conversation ends in a decision table or summary, that is authoritative wherever it conflicts with earlier prose.",
+  "The input may arrive as numbered compressed slices followed by verbatim recent turns. Slice numbers are chronological: a later slice overrules an earlier one.",
+  "Decisions the agent made that the user did not contradict are worth reporting, but user silence only ever yields tentative status.",
+  "Do NOT include the agent's self-assessment or any claim that the work is correct or complete \u2014 the reviewer must judge that independently.",
+  "Every statement must trace to something in the conversation. If a heading has nothing real to report, say so rather than inventing plausible-sounding content.",
+  "Write concise markdown bullets under those numbered headings, under 500 words total, in the same language as the conversation. Reply with the brief only."
 ].join("\n");
-function capText(text2, max) {
-  return text2.length > max ? text2.slice(0, max) + "\u2026" : text2;
+var REVERSAL_SYSTEM_PROMPT = [
+  "You read a coding-agent conversation and report ONLY the conclusions that were later reversed, dropped, or replaced.",
+  "A design conversation converges by killing its own hypotheses: an option explored in depth early is often rejected later in one sentence, so weigh finality over how much was written.",
+  "A reversal counts whether the user or the agent made it. Ignore rewording that leaves the decision intact.",
+  "Output one line per reversal: SUPERSEDED: <what was concluded earlier> -> <what replaced it> (<why>)",
+  "Report the most decisive reversals first and stay under 200 words \u2014 a list that runs long will be discarded whole, so prioritise rather than pad.",
+  "If nothing was reversed, reply with exactly: NONE"
+].join("\n");
+var COMPACT_SYSTEM_PROMPT = [
+  "You compress one slice of a coding-agent conversation. A later pass will distill the whole thing into an intent brief for a code reviewer.",
+  "You are told which slice of how many this is. Other slices are compressed separately and you cannot see them, so never present a conclusion here as final \u2014 a later slice may overturn it.",
+  "Preserve, in this order of priority:",
+  "1. Reversals and corrections \u2014 any conclusion that was dropped, replaced, or corrected. Quote the deciding sentence verbatim.",
+  "2. Constraints and decisions the user stated, and approaches that were rejected, with the reason. Preserve who made each decision \u2014 an explicit user confirmation must stay distinguishable from an agent proposal the user never answered; quote the user's deciding sentence when there is one.",
+  "3. Behaviour established by actually running something: observed output, exit codes, error text, which stream an error arrived on.",
+  "4. Trade-offs that were acknowledged and accepted.",
+  "Drop: exploration narrative, restated background, tool mechanics, and anything a reviewer could recover by reading the code.",
+  "Stay under 800 words. An over-long slice is discarded whole rather than cut short, so drop lower-priority material instead of running over.",
+  "Keep the conversation's own language. Reply with the compressed slice only, no preamble."
+].join("\n");
+function timeoutForInput(chars) {
+  return Math.min(MAX_TIMEOUT_MS, BASE_TIMEOUT_MS + Math.floor(chars / 1e3) * TIMEOUT_MS_PER_1K_CHARS);
+}
+function isSizeRelatedFailure(error48) {
+  const err = error48;
+  if (err?.name === "AbortError" || err?.name === "TimeoutError") return true;
+  const message = err?.message ?? "";
+  return /context[ _-]?length|maximum context|too many tokens|token limit|prompt is too long|request too large|timed? ?out/i.test(
+    message
+  );
 }
-function serializeConversationForBrief(messages) {
-  const lines = [];
+function withinLimit(text2, max, label) {
+  if (text2.length <= max) return true;
+  console.warn(`[ReviewBrief] ${label} ran to ${text2.length} chars (limit ${max}) \u2014 discarding rather than truncating`);
+  return false;
+}
+function capMessage(text2) {
+  return text2.length > PER_MESSAGE_MAX_CHARS ? text2.slice(0, PER_MESSAGE_MAX_CHARS) + "\u2026" : text2;
+}
+function toEntries(messages) {
+  const entries = [];
   for (const msg of messages) {
     if (!msg) continue;
     if (msg.type === "user" && !msg.event) {
       const text2 = extractUserText(msg.content).trim();
-      if (text2) lines.push(`User: ${capText(text2, PER_MESSAGE_MAX_CHARS)}`);
+      if (text2) entries.push({ role: "User", text: capMessage(text2) });
     } else if (msg.type === "assistant" && typeof msg.content === "string") {
       const text2 = msg.content.trim();
-      if (text2) lines.push(`Agent: ${capText(text2, PER_MESSAGE_MAX_CHARS)}`);
+      if (text2) entries.push({ role: "Agent", text: capMessage(text2) });
     }
   }
-  const full = lines.join("\n\n");
-  if (full.length <= TOTAL_INPUT_MAX_CHARS) return full;
-  return `${full.slice(0, HEAD_CHARS)}
-
-[\u2026 middle of the conversation omitted \u2026]
-
-${full.slice(-TAIL_CHARS)}`;
+  return entries;
 }
-async function generateIntentBriefWithModel(model, conversation, options = {}) {
-  if (conversation.trim().length === 0) return null;
-  const telemetry = options.userId ? {
-    isEnabled: true,
-    functionId: "review-intent-brief",
-    metadata: {
-      userId: options.userId,
-      tags: ["vibedeckx", "review-intent-brief"]
+function render(entry) {
+  return `${entry.role}: ${entry.text}`;
+}
+function renderAll(entries) {
+  return entries.map(render).join(SEP);
+}
+function serializeConversationForBrief(messages) {
+  return renderAll(toEntries(messages));
+}
+function splitRecent(entries, maxChars) {
+  let size = 0;
+  let i = entries.length;
+  while (i > 0) {
+    const len = render(entries[i - 1]).length + SEP.length;
+    if (size + len > maxChars) break;
+    size += len;
+    i--;
+  }
+  return { older: entries.slice(0, i), recent: entries.slice(i) };
+}
+function chunkEntries(entries, maxChars) {
+  const chunks = [];
+  let current = [];
+  let size = 0;
+  for (const entry of entries) {
+    const len = render(entry).length + SEP.length;
+    if (current.length > 0 && size + len > maxChars) {
+      chunks.push(current);
+      current = [];
+      size = 0;
     }
+    current.push(entry);
+    size += len;
+  }
+  if (current.length > 0) chunks.push(current);
+  return chunks;
+}
+function telemetryFor(functionId, userId) {
+  return userId ? {
+    isEnabled: true,
+    functionId,
+    metadata: { userId, tags: ["vibedeckx", functionId] }
   } : void 0;
+}
+async function compactChunk(model, chunk, part, total, userId) {
   try {
     const result = await generateText({
       model,
-      system: SYSTEM_PROMPT2,
-      prompt: `Distill this conversation into an intent brief:
+      system: COMPACT_SYSTEM_PROMPT,
+      prompt: `Slice ${part} of ${total}.
+
+${chunk}`,
+      timeout: timeoutForInput(chunk.length),
+      maxOutputTokens: COMPACT_SUMMARY_MAX_TOKENS,
+      experimental_telemetry: telemetryFor("review-intent-brief-compact", userId)
+    });
+    const { text: rawText, finishReason } = result;
+    if (finishReason === "length") {
+      console.warn(`[ReviewBrief] slice ${part}/${total} summary hit the token cap \u2014 discarding`);
+      return null;
+    }
+    const text2 = (rawText ?? "").trim();
+    if (text2.length === 0) {
+      console.warn(`[ReviewBrief] slice ${part}/${total} came back empty (finishReason: ${finishReason ?? "unknown"})`);
+      return null;
+    }
+    return withinLimit(text2, COMPACT_SUMMARY_MAX_CHARS, `slice ${part}/${total} summary`) ? text2 : null;
+  } catch (error48) {
+    if (isSizeRelatedFailure(error48)) throw error48;
+    console.warn(`[ReviewBrief] compaction of slice ${part}/${total} failed:`, error48.message);
+    return null;
+  }
+}
+async function compactConversation(model, messages, options = {}) {
+  const sliceChars = options.sliceChars ?? COMPACT_SLICE_BUDGETS[0];
+  const recentChars = Math.max(MIN_RECENT_VERBATIM_CHARS, Math.floor(sliceChars * RECENT_VERBATIM_RATIO));
+  const entries = toEntries(messages);
+  if (entries.length === 0) return null;
+  const { older, recent } = splitRecent(entries, recentChars);
+  if (older.length === 0) return null;
+  const chunks = chunkEntries(older, sliceChars);
+  if (chunks.length > MAX_FOLD_CALLS) {
+    console.warn(
+      `[ReviewBrief] conversation needs ${chunks.length} slices of ${sliceChars} chars (max ${MAX_FOLD_CALLS}) \u2014 falling back to tier 2`
+    );
+    return null;
+  }
+  const settled = await Promise.allSettled(
+    chunks.map((chunk, i) => compactChunk(model, renderAll(chunk), i + 1, chunks.length, options.userId))
+  );
+  const sizeRejection = settled.find((r) => r.status === "rejected" && isSizeRelatedFailure(r.reason));
+  if (sizeRejection) throw sizeRejection.reason;
+  if (settled.some((r) => r.status === "rejected" || r.value === null)) return null;
+  const parts = settled.map(
+    (r, i) => `[Compressed slice ${i + 1} of ${chunks.length}]
+${r.value}`
+  );
+  if (recent.length > 0) parts.push(`[Recent turns, verbatim]
+${renderAll(recent)}`);
+  return parts.join(SEP);
+}
+function buildBriefPrompt(conversation, reversals) {
+  const preamble = reversals ? "Conclusions this conversation reversed. The right-hand side is the current state \u2014 never report the left-hand side as what was decided:\n" + reversals + SEP : "";
+  return `${preamble}Distill this conversation into an intent brief:
+
+${conversation}`;
+}
+function appendCompactionNote(brief, compacted) {
+  return compacted ? `${brief}${SEP}${COMPACTION_NOTE}` : brief;
+}
+async function extractReversalsWithModel(model, conversation, options = {}) {
+  if (conversation.trim().length < REVERSAL_MIN_CHARS) return null;
+  try {
+    const result = await generateText({
+      model,
+      system: REVERSAL_SYSTEM_PROMPT,
+      prompt: `Report the reversals in this conversation:
 
 ${conversation}`,
-      timeout: AI_TIMEOUT_MS2,
-      experimental_telemetry: telemetry
+      timeout: JUDGMENT_TIMEOUT_MS,
+      maxOutputTokens: REVERSALS_MAX_TOKENS,
+      experimental_telemetry: telemetryFor("review-intent-brief-reversals", options.userId)
     });
-    const text2 = (result.text ?? "").trim();
-    return text2.length > 0 ? capText(text2, BRIEF_MAX_CHARS) : null;
+    const { text: rawText, finishReason } = result;
+    if (finishReason === "length") {
+      console.warn("[ReviewBrief] reversal list hit the token cap \u2014 discarding");
+      return null;
+    }
+    const text2 = (rawText ?? "").trim();
+    if (text2.length === 0 || /^none\b/i.test(text2)) return null;
+    return withinLimit(text2, REVERSALS_MAX_CHARS, "reversal list") ? text2 : null;
+  } catch (error48) {
+    if (!isSizeRelatedFailure(error48)) throw error48;
+    console.warn("[ReviewBrief] reversal pass hit a size limit \u2014 skipping it:", error48.message);
+    return null;
+  }
+}
+async function generateIntentBriefWithModel(model, conversation, options = {}) {
+  if (conversation.trim().length === 0) return null;
+  try {
+    const result = await generateText({
+      model,
+      system: SYSTEM_PROMPT2,
+      prompt: buildBriefPrompt(conversation, options.reversals ?? null),
+      timeout: JUDGMENT_TIMEOUT_MS,
+      maxOutputTokens: BRIEF_MAX_TOKENS,
+      experimental_telemetry: telemetryFor("review-intent-brief", options.userId)
+    });
+    const { text: rawText, finishReason } = result;
+    const text2 = (rawText ?? "").trim();
+    if (text2.length === 0) {
+      console.warn(`[ReviewBrief] model returned empty text (finishReason: ${finishReason ?? "unknown"})`);
+      return null;
+    }
+    if (finishReason === "length") return text2.slice(0, BRIEF_MAX_CHARS) + BRIEF_TRUNCATION_NOTE;
+    return text2.length <= BRIEF_MAX_CHARS ? text2 : text2.slice(0, BRIEF_MAX_CHARS) + BRIEF_TRUNCATION_NOTE;
   } catch (error48) {
+    if (options.rethrowSizeFailures && isSizeRelatedFailure(error48)) throw error48;
     console.warn("[ReviewBrief] AI generation failed:", error48.message);
     return null;
   }
 }
+async function distill(model, conversation, userId, rethrowSizeFailures) {
+  const reversals = await extractReversalsWithModel(model, conversation, { userId });
+  return generateIntentBriefWithModel(model, conversation, { userId, reversals, rethrowSizeFailures });
+}
 async function generateIntentBrief(storage, userId, messages) {
   try {
-    if (!await isChatModelConfigured(storage, userId)) return null;
+    const config2 = await getChatProviderConfig(storage, userId);
+    const mainOk = isModelConfigured(config2, config2.main);
+    const fastOk = isModelConfigured(config2, config2.fast);
+    if (!mainOk && !fastOk) return null;
     const conversation = serializeConversationForBrief(messages);
     if (!conversation) return null;
-    return await generateIntentBriefWithModel(await resolveFastChatModel(storage, userId), conversation, { userId });
+    const distillModel = mainOk ? await resolveChatModel(storage, userId) : await resolveFastChatModel(storage, userId);
+    const compactModel = fastOk ? await resolveFastChatModel(storage, userId) : distillModel;
+    if (conversation.length <= COMPACT_TRIGGER_CHARS) {
+      try {
+        const brief = await distill(distillModel, conversation, userId, true);
+        return brief === null ? null : appendCompactionNote(brief, false);
+      } catch (error48) {
+        if (!isSizeRelatedFailure(error48)) throw error48;
+        console.warn("[ReviewBrief] full conversation rejected as too large \u2014 compacting and retrying");
+      }
+    }
+    for (const sliceChars of COMPACT_SLICE_BUDGETS) {
+      try {
+        const compacted = await compactConversation(compactModel, messages, { userId, sliceChars });
+        if (compacted === null) return null;
+        const brief = await distill(distillModel, compacted, userId, true);
+        return brief === null ? null : appendCompactionNote(brief, true);
+      } catch (error48) {
+        if (!isSizeRelatedFailure(error48)) throw error48;
+        console.warn(`[ReviewBrief] ${sliceChars}-char slices still too large \u2014 retrying smaller`);
+      }
+    }
+    console.warn("[ReviewBrief] could not compact the conversation small enough \u2014 falling back to tier 2");
+    return null;
   } catch (error48) {
     console.warn("[ReviewBrief] generation failed:", error48.message);
     return null;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vibedeckx/darwin-arm64",
-  "version": "0.2.16",
+  "version": "0.2.18",
   "description": "Vibedeckx platform binaries for macOS ARM64",
   "os": [
     "darwin"