onbuzz 4.8.2 → 4.9.0

package/src/tools/taskManagerTool.js

@@ -345,6 +345,57 @@ Always use a detailed task description to provide context for the task, and leve
     return description;
   }
 
+  /**
+   * Build the `currentFocus` signal emitted on every taskmanager
+   * response. Tells the agent which task it is CURRENTLY working on
+   * and warns about discipline violations (zero or multiple
+   * in-progress tasks).
+   *
+   * Three shapes:
+   *   { current: {taskId,title,description}, inProgressCount: 1, ok: true }
+   *   { current: null, inProgressCount: 0, ok: false,
+   *     warning: 'No task is in_progress…' }
+   *   { current: [...], inProgressCount: N>1, ok: false,
+   *     warning: 'Multiple tasks are in_progress…' }
+   *
+   * Production rationale: agents lose task focus after cascading tool
+   * errors and revert to earlier tasks. Surfacing the focus on every
+   * taskmanager turn keeps the anchor visible. See baseTool.js
+   * OPERATING POSTURE → TASK DISCIPLINE section.
+   *
+   * @private
+   */
+  _buildCurrentFocus(tasks) {
+    const inProgress = (tasks || []).filter(t => t.status === 'in_progress');
+    const shortDesc = (d) => {
+      if (typeof d !== 'string') return '';
+      const trimmed = d.trim();
+      return trimmed.length > 140 ? trimmed.slice(0, 140) + '…' : trimmed;
+    };
+    if (inProgress.length === 1) {
+      const t = inProgress[0];
+      return {
+        ok: true,
+        inProgressCount: 1,
+        current: { taskId: t.id, title: t.title, description: shortDesc(t.description) },
+      };
+    }
+    if (inProgress.length === 0) {
+      return {
+        ok: false,
+        inProgressCount: 0,
+        current: null,
+        warning: 'No task is in_progress. BEFORE doing any work, call `taskmanager` → `update` with status `in_progress` on the task you are about to work on. The transition IS the signal that you have started.',
+      };
+    }
+    return {
+      ok: false,
+      inProgressCount: inProgress.length,
+      current: inProgress.map(t => ({ taskId: t.id, title: t.title, description: shortDesc(t.description) })),
+      warning: `Multiple tasks are in_progress (${inProgress.length}). Pick ONE to remain in_progress; transition the others to \`blocked\` or \`completed\` so the task list reflects what you are actually doing right now.`,
+    };
+  }
+
   /**
    * Extract memory IDs from task descriptions
    * @private
@@ -436,12 +487,15 @@ Always use a detailed task description to provide context for the task, and leve
           firstAction = unwrappedAction;
         }
 
-        // Extract action type from 'type' field in the actions array
-        if (firstAction.type && !params.action) {
-          params.action = firstAction.type;
+        // Extract action type — accept both `type` (canonical) and
+        // `action` (used by jobdone/docx/excel/pdf). Cross-tool field-
+        // name unification.
+        const itemType = firstAction.type || firstAction.action;
+        if (itemType && !params.action) {
+          params.action = itemType;
           // Merge all other properties from the action object into params
           for (const [key, value] of Object.entries(firstAction)) {
-            if (key !== 'type' && !(key in params)) {
+            if (key !== 'type' && key !== 'action' && !(key in params)) {
               params[key] = value;
             }
           }
@@ -506,7 +560,7 @@ Always use a detailed task description to provide context for the task, and leve
           result = await this.createTask(agent, params, context);
           break;
         case 'update':
-          result = await this.updateTask(agent, params);
+          result = await this.updateTask(agent, params, context);
           break;
         case 'list':
           result = await this.listTasks(agent, params);
@@ -583,16 +637,22 @@ Always use a detailed task description to provide context for the task, and leve
       // so UI views (TaskPanel, TeamTaskBoard) can render the agent's current
       // state without fetching on every render. Previously only the `list`
       // action's inner result carried `tasks`, so every `create`/`update`/
-      // `complete`/`cancel`/`clear` call left the panel looking stale: the
-      // renderer's walk-back found the latest taskmanager call but found no
-      // tasks in its result and showed an empty panel. Emitting the full
-      // list on every action makes the most-recent-taskmanager-result
-      // contract always authoritative.
+      // `complete`/`cancel`/`clear` call left the panel looking stale.
+      //
+      // `currentFocus` is the task-discipline signal — see baseTool.js
+      // OPERATING POSTURE addition. It tells the agent which task it is
+      // CURRENTLY working on, warns when zero or multiple tasks are
+      // in_progress, and stays visible across every taskmanager turn so
+      // tool-error cascades don't shake the agent off its task. Production
+      // observation: Kimi-K2.6 + gpt-5.4 lost task focus mid-session after
+      // cascading tool failures and reverted to earlier tasks. Surfacing
+      // the focus on every response keeps the anchor visible.
       return {
         success: true,
         action,
         result,
         tasks: agent.taskList.tasks,
+        currentFocus: this._buildCurrentFocus(agent.taskList.tasks),
         summary: this.generateTaskSummary(agent.taskList) + noTasksHint
       };
 
@@ -858,7 +918,7 @@ Always use a detailed task description to provide context for the task, and leve
    * Update an existing task
    * @private
    */
-  async updateTask(agent, params) {
+  async updateTask(agent, params, context = {}) {
     const { taskId, status, priority, title, description } = params;
 
     if (!taskId) {
@@ -870,6 +930,8 @@ Always use a detailed task description to provide context for the task, and leve
       throw new Error(`Task not found: ${taskId}`);
     }
 
+    const prevStatus = task.status;
+
     if (status) {
       if (!this.taskStatuses.includes(status.toLowerCase())) {
         throw new Error(`Invalid status: ${status}. Must be: ${this.taskStatuses.join(', ')}`);
@@ -892,13 +954,86 @@ Always use a detailed task description to provide context for the task, and leve
     // only mutate fields the model actually set to a meaningful value.
     if (typeof title === 'string' && title.trim().length > 0) task.title = title;
     if (typeof description === 'string' && description.trim().length > 0) task.description = description;
-    
+
     task.updatedAt = new Date().toISOString();
 
-    return {
+    const result = {
       message: 'Task updated successfully',
-      task
+      task,
     };
+
+    // ── Task-transition memory handler ────────────────────────────────
+    // When a task transitions INTO `in_progress`, we either:
+    //   (a) Auto-load its attached memory (if the description carries a
+    //       `[memory:mem-...]` marker) and embed the content in the tool
+    //       result so the agent reads it on the same turn — no extra
+    //       `memory read` call required, no nudge needed.
+    //   (b) If no memory is attached, emit a one-line nudge pointing the
+    //       agent at the `## AGENT MEMORY` catalog in the system prompt
+    //       so it can scan for adjacent context.
+    //
+    // Production observation (Monopoly trace): 0 memory ops across
+    // 1556 messages. Memories that DO get created are not consulted.
+    // Surfacing the attached memory at the transition moment makes it
+    // un-missable; the conditional nudge keeps the prompt clean when
+    // there's nothing to attach.
+    const becameInProgress = prevStatus !== 'in_progress' && task.status === 'in_progress';
+    if (becameInProgress && context.agentId) {
+      const memMatch = typeof task.description === 'string'
+        ? task.description.match(/\[memory:(mem-[A-Za-z0-9_-]+)\]/)
+        : null;
+      if (memMatch) {
+        try {
+          const memoryService = getMemoryService(this.logger);
+          await memoryService.initialize();
+          const all = await memoryService.loadMemories(context.agentId);
+          const mem = (all || []).find(m => m.id === memMatch[1]);
+          if (mem) {
+            result.attachedMemory = {
+              id: mem.id,
+              title: mem.title,
+              content: mem.content,
+              _hint: 'This memory was auto-loaded because the task references it. Use `memory append` (id: ' + mem.id + ') if your work yields new findings or methodology worth recording.',
+            };
+          } else {
+            // Reference exists but memory was deleted — surface that fact.
+            result.attachedMemory = {
+              id: memMatch[1],
+              error: 'Memory referenced in task description was not found (may have been deleted). Consider attaching a fresh one.',
+            };
+          }
+        } catch (e) {
+          this.logger?.debug?.('Auto-load of attached memory failed', { error: e.message });
+        }
+      } else {
+        // No memory attached — emit a nudge pointing at the catalog.
+        //
+        // Wording is gated by LOXIA_MEMORY_NUDGE_VARIANT (env) so the
+        // memory-bench can A/B baseline ('descriptive') vs directive
+        // ('directive') on production code WITHOUT requiring two
+        // separate builds.
+        //
+        // Default is 'directive' (the ship in 9a6957f). Validation
+        // pass scheduled to compare against 'descriptive' on
+        // production-relevant models (Kimi-K2.6, gpt-5.4) before
+        // confirming the ship or reverting.
+        const variant = (process.env.LOXIA_MEMORY_NUDGE_VARIANT || 'directive').toLowerCase();
+        if (variant === 'descriptive' || variant === 'baseline') {
+          result.memoryNudge =
+            'No memory is attached to this task (description has no `[memory:mem-...]` marker). ' +
+            'Scan the `## AGENT MEMORY` catalog in your system prompt for adjacent context that may inform this work. ' +
+            'If you produce findings or methodology worth preserving, create a memory with `memory add` (title prefix `task-context/...` recommended) and reference it in this task.';
+        } else {
+          // 'directive' (default / current ship).
+          result.memoryNudge =
+            'No memory is attached to this task (description has no `[memory:mem-...]` marker). ' +
+            'Your VERY NEXT tool call MUST be `memory list` to check for applicable rules, preferences, or prior findings before doing the work. ' +
+            'If you produce findings or methodology worth preserving, create a memory with `memory add` (title prefix `task-context/...` recommended) and reference it in this task.';
+        }
+      }
+    }
+
+    return result;
   }
 
   /**

package/src/tools/terminalTool.js

@@ -380,11 +380,14 @@ After each invocation you will get a tool-result embedded in the following user
     } else {
       // Validate each action
       for (const [index, action] of params.actions.entries()) {
+        // Accept both `type` and `action` as item discriminator
+        // (cross-tool field-name unification).
+        if (!action.type && action.action) action.type = action.action;
         if (!action.type) {
           errors.push(`Action ${index + 1}: type is required`);
           continue;
         }
-        
+
         switch (action.type) {
           case 'run-command':
             if (!action.command || !action.command.trim()) {
@@ -431,6 +434,25 @@ After each invocation you will get a tool-result embedded in the following user
    * @returns {Promise<Object>} Execution result
    */
   async execute(params, context) {
+    // Forgiving normalization — production agents commonly call with
+    // the natural singular shape `{command: "ls"}` instead of the
+    // documented plural `{actions: [{type: "run-command", command: "ls"}]}`.
+    // Live trace: gpt-5.4 + Kimi-K2.6 both hit "actions is not iterable"
+    // on this. Accept singular forms here and promote to canonical.
+    if ((!params.actions || !Array.isArray(params.actions)) && params.command) {
+      const { command, type, ...rest } = params;
+      params = { ...params, actions: [{ type: type || 'run-command', command, ...rest }] };
+    }
+
+    // Normalize discriminator on each action item (accept both `type`
+    // and `action`). Same rationale as fileSystemTool — execute() may
+    // be called bypassing validation in production.
+    if (Array.isArray(params.actions)) {
+      for (const a of params.actions) {
+        if (a && !a.type && a.action) a.type = a.action;
+      }
+    }
+
     const { actions, timeout: customTimeout, async: forceAsync } = params;
     const { agentId, projectDir, directoryAccess } = context;
 

package/src/tools/visionTool.js

@@ -123,6 +123,13 @@ ANSWERS: Structured, direct answers only — no filler text.`;
   }
 
   async execute(params, context = {}) {
+    // Reverse-forgiveness: accept plural envelope {actions:[{type/action}]}
+    // even though this tool's canonical shape is singular.
+    if (Array.isArray(params?.actions) && params.actions.length > 0) {
+      const a = params.actions[0] || {};
+      params = { ...params, ...a, action: a.type || a.action || params.action };
+    }
+
     this._ensureCleanup();
 
     try {

package/src/tools/visualEditorTool.js

@@ -319,6 +319,13 @@ Code:
    * @returns {Promise<Object>} Execution result
    */
   async execute(params, context = {}) {
+    // Reverse-forgiveness: accept plural {actions:[{...}]} envelope
+    // (cross-tool unification).
+    if (Array.isArray(params?.actions) && params.actions.length > 0) {
+      const a = params.actions[0] || {};
+      params = { ...params, ...a, action: a.type || a.action || params.action };
+    }
+
     // FIRST THING: an unconditional console.log so we can always see that
     // the tool was invoked — before any early-return (parse failure,
     // bridge disabled, validation fail) or thrown error could swallow it.

package/src/tools/whatsappTool.js

@@ -242,6 +242,13 @@ Check connection status:
   }
 
   async execute(params, context) {
+    // Reverse-forgiveness: accept plural {actions:[{...}]} envelope
+    // (cross-tool unification).
+    if (Array.isArray(params?.actions) && params.actions.length > 0) {
+      const a = params.actions[0] || {};
+      params = { ...params, ...a, action: a.type || a.action || params.action };
+    }
+
     // Unwrap TagParser objects
     const unwrapped = {};
     for (const [key, value] of Object.entries(params)) {

package/src/utilities/authCache.js

@@ -8,7 +8,7 @@
  * testable without standing up Express.
  *
  * Cache shape:
- *   { user, hasApiKey, jwt, timestamp }
+ *   { user, hasApiKey, jwt, timestamp, expiresAt }
  *
  *   - user       — { id, email, ... } or null
  *   - hasApiKey  — bool
@@ -17,6 +17,10 @@
  *                  service verifies JWTs directly, not lx_* keys).
  *   - timestamp  — ISO of when this cache row was written; used for
  *                  staleness checks via `?since=` polling.
+ *   - expiresAt  — ISO after which the session is considered expired.
+ *                  Checked by `projectAuthStatus` and by the account-router
+ *                  JWT guard. Default TTL: 24 h (configurable via
+ *                  `auth.sessionTtlHours`).
  *
  * Bug history this guards against:
  *   - 2026-05: marketplace publish 401 "Authentication required" because
@@ -27,6 +31,14 @@
  *     marketplace calls → 401.
  */
 
+const DEFAULT_SESSION_TTL_HOURS = 24;
+const MS_PER_HOUR = 60 * 60 * 1000;
+
+function computeExpiresAt(ttlHours = DEFAULT_SESSION_TTL_HOURS) {
+  const hours = typeof ttlHours === 'number' && ttlHours > 0 ? ttlHours : DEFAULT_SESSION_TTL_HOURS;
+  return new Date(Date.now() + hours * MS_PER_HOUR).toISOString();
+}
+
 /**
  * Build a cache row for a successful OAuth callback. The JWT is the
  * critical field — without it, the polling fallback can't restore
@@ -37,14 +49,16 @@
  * @param {string|null} input.jwt  - Raw JWT from the OAuth callback
  * @param {boolean} input.hasApiKey
  * @param {string} [input.timestamp] - Override (for tests); defaults to now
+ * @param {number} [input.ttlHours] - Session TTL in hours; defaults to 24
  * @returns {Object} cache row
  */
-export function recordOAuthResult({ user, jwt, hasApiKey, timestamp } = {}) {
+export function recordOAuthResult({ user, jwt, hasApiKey, timestamp, ttlHours } = {}) {
   return {
     user: user || null,
     hasApiKey: !!hasApiKey,
     jwt: typeof jwt === 'string' && jwt.length > 0 ? jwt : null,
     timestamp: timestamp || new Date().toISOString(),
+    expiresAt: computeExpiresAt(ttlHours),
   };
 }
 
@@ -59,9 +73,10 @@ export function recordOAuthResult({ user, jwt, hasApiKey, timestamp } = {}) {
  * @param {Object} input.user
  * @param {Object|null} input.prevCache - previous lastAuthResult
  * @param {string} [input.timestamp]
+ * @param {number} [input.ttlHours] - Session TTL in hours; defaults to 24
  * @returns {Object} cache row
  */
-export function recordResolvedUser({ user, prevCache, timestamp } = {}) {
+export function recordResolvedUser({ user, prevCache, timestamp, ttlHours } = {}) {
   const carriedJwt = prevCache && typeof prevCache.jwt === 'string' && prevCache.jwt.length > 0
     ? prevCache.jwt
     : null;
@@ -70,9 +85,24 @@ export function recordResolvedUser({ user, prevCache, timestamp } = {}) {
     hasApiKey: true,                        // resolve path requires an API key
     jwt: carriedJwt,
     timestamp: timestamp || new Date().toISOString(),
+    expiresAt: computeExpiresAt(ttlHours),
   };
 }
 
+/**
+ * Returns true when the cache row carries an `expiresAt` that has
+ * already passed (or is exactly now). Rows without `expiresAt`
+ * (legacy cache from before this field was added) are treated as
+ * expired so the user is forced to re-auth once.
+ *
+ * @param {Object|null} cache
+ * @returns {boolean}
+ */
+export function isAuthExpired(cache) {
+  if (!cache || typeof cache.expiresAt !== 'string') return true;
+  return new Date(cache.expiresAt).getTime() <= Date.now();
+}
+
 /**
  * Project a cache row into the JSON the `/api/auth/status` endpoint
  * returns. Implements two pieces of business logic:
@@ -88,20 +118,30 @@ export function recordResolvedUser({ user, prevCache, timestamp } = {}) {
  *      pair is always coherent so the FE never sees user=null with
  *      jwt=<something>.
  *
+ *   3. Expiry gate: if `expiresAt` has passed, the user is reported
+ *      as null, forcing the frontend to re-authenticate.
+ *
  * @param {Object|null} cache - lastAuthResult value
  * @param {Object} [options]
  * @param {string} [options.since] - ISO timestamp from the caller
  * @param {boolean} [options.hasApiKey] - hasApiKey value at request time
  *   (cache.hasApiKey is the value AT CACHE TIME — the live answer comes
  *   from the apiKeyManager). The route passes the live one in.
- * @returns {{ success: true, authenticated: boolean, user: Object|null, jwt: string|null, timestamp: string|null }}
+ * @returns {{ success: true, authenticated: boolean, user: Object|null, jwt: string|null, timestamp: string|null, expired: boolean }}
  */
 export function projectAuthStatus(cache, options = {}) {
   const since = options.since;
   const liveHasApiKey = options.hasApiKey === undefined ? !!cache?.hasApiKey : !!options.hasApiKey;
 
   let user = cache?.user || null;
-  if (since && cache?.timestamp) {
+
+  // Expiry gate — treat expired sessions as unauthenticated.
+  const expired = isAuthExpired(cache);
+  if (expired) {
+    user = null;
+  }
+
+  if (!expired && since && cache?.timestamp) {
     if (new Date(cache.timestamp).getTime() <= new Date(since).getTime()) {
       // Cache predates the login attempt the caller is polling for.
       user = null;
@@ -113,9 +153,10 @@ export function projectAuthStatus(cache, options = {}) {
 
   return {
     success: true,
-    authenticated: liveHasApiKey,
+    authenticated: liveHasApiKey && !expired,
     user,
     jwt,
     timestamp: cache?.timestamp || null,
+    expired,
   };
 }

package/src/utilities/authCache.js.backup-1779570472481

@@ -0,0 +1,121 @@
+/**
+ * Auth-result cache shape + transitions.
+ *
+ * The webServer caches the most recent auth event in `lastAuthResult` so
+ * a frontend that missed the WS broadcast (page reload, dropped
+ * connection, race) can recover via polling `/api/auth/status`. The
+ * cache shape and three state transitions live here so they're
+ * testable without standing up Express.
+ *
+ * Cache shape:
+ *   { user, hasApiKey, jwt, timestamp }
+ *
+ *   - user       — { id, email, ... } or null
+ *   - hasApiKey  — bool
+ *   - jwt        — raw bearer token from OAuth, or null. Mandatory for
+ *                  marketplace publish/install/rate calls (the marketplace
+ *                  service verifies JWTs directly, not lx_* keys).
+ *   - timestamp  — ISO of when this cache row was written; used for
+ *                  staleness checks via `?since=` polling.
+ *
+ * Bug history this guards against:
+ *   - 2026-05: marketplace publish 401 "Authentication required" because
+ *     OAuth callback cached only {user, hasApiKey, timestamp} without
+ *     the JWT, then a page reload triggered the polling fallback which
+ *     returned the user but no JWT. The FE thought "I'm signed in" but
+ *     localStorage.loxia-auth-token stayed empty → no Bearer header on
+ *     marketplace calls → 401.
+ */
+
+/**
+ * Build a cache row for a successful OAuth callback. The JWT is the
+ * critical field — without it, the polling fallback can't restore
+ * marketplace auth on a refresh.
+ *
+ * @param {Object} input
+ * @param {Object} input.user      - User account object
+ * @param {string|null} input.jwt  - Raw JWT from the OAuth callback
+ * @param {boolean} input.hasApiKey
+ * @param {string} [input.timestamp] - Override (for tests); defaults to now
+ * @returns {Object} cache row
+ */
+export function recordOAuthResult({ user, jwt, hasApiKey, timestamp } = {}) {
+  return {
+    user: user || null,
+    hasApiKey: !!hasApiKey,
+    jwt: typeof jwt === 'string' && jwt.length > 0 ? jwt : null,
+    timestamp: timestamp || new Date().toISOString(),
+  };
+}
+
+/**
+ * Build a cache row when a user is resolved from an API key alone (no
+ * OAuth event, so no JWT minted in this flow). PRESERVES any JWT from
+ * a previous OAuth login — overwriting `lastAuthResult` with no `jwt`
+ * field would silently wipe marketplace auth for users who hit
+ * /auth/resolve after /auth/callback.
+ *
+ * @param {Object} input
+ * @param {Object} input.user
+ * @param {Object|null} input.prevCache - previous lastAuthResult
+ * @param {string} [input.timestamp]
+ * @returns {Object} cache row
+ */
+export function recordResolvedUser({ user, prevCache, timestamp } = {}) {
+  const carriedJwt = prevCache && typeof prevCache.jwt === 'string' && prevCache.jwt.length > 0
+    ? prevCache.jwt
+    : null;
+  return {
+    user: user || null,
+    hasApiKey: true,                        // resolve path requires an API key
+    jwt: carriedJwt,
+    timestamp: timestamp || new Date().toISOString(),
+  };
+}
+
+/**
+ * Project a cache row into the JSON the `/api/auth/status` endpoint
+ * returns. Implements two pieces of business logic:
+ *
+ *   1. Staleness gate via `?since=`: if the caller is polling for an
+ *      auth event AFTER a known timestamp, but the cache row predates
+ *      that timestamp, the row is considered "not the auth we're
+ *      waiting for" and the user is reported as null. Prevents a
+ *      previously-cached login from masquerading as the in-flight one.
+ *
+ *   2. JWT only flows out when a fresh user is reported. When `user`
+ *      is null (no cached row, or stale), `jwt` is also null — the
+ *      pair is always coherent so the FE never sees user=null with
+ *      jwt=<something>.
+ *
+ * @param {Object|null} cache - lastAuthResult value
+ * @param {Object} [options]
+ * @param {string} [options.since] - ISO timestamp from the caller
+ * @param {boolean} [options.hasApiKey] - hasApiKey value at request time
+ *   (cache.hasApiKey is the value AT CACHE TIME — the live answer comes
+ *   from the apiKeyManager). The route passes the live one in.
+ * @returns {{ success: true, authenticated: boolean, user: Object|null, jwt: string|null, timestamp: string|null }}
+ */
+export function projectAuthStatus(cache, options = {}) {
+  const since = options.since;
+  const liveHasApiKey = options.hasApiKey === undefined ? !!cache?.hasApiKey : !!options.hasApiKey;
+
+  let user = cache?.user || null;
+  if (since && cache?.timestamp) {
+    if (new Date(cache.timestamp).getTime() <= new Date(since).getTime()) {
+      // Cache predates the login attempt the caller is polling for.
+      user = null;
+    }
+  }
+
+  // Pair invariant: jwt is null whenever user is null.
+  const jwt = user ? (cache?.jwt || null) : null;
+
+  return {
+    success: true,
+    authenticated: liveHasApiKey,
+    user,
+    jwt,
+    timestamp: cache?.timestamp || null,
+  };
+}

package/src/utilities/constants.js

@@ -656,22 +656,30 @@ const COMPACTION_CONFIG = {
   MIN_MIDDLE_SEGMENT_PERCENTAGE: 0.50, // Middle must be at least 50% of messages
   MAX_BOOKEND_PERCENTAGE: 0.50, // Beginning + end together capped at 50%
 
-  // Recommended model pool for compaction (validated against live model catalog at runtime)
-  // Names MUST match catalog keys exactly (no azure-openai- prefix)
-  // Ordered by context window size (largest first) to handle very large conversations
+  // Recommended model pool for compaction (validated against live model catalog at runtime).
+  // Names MUST match catalog keys exactly (no azure-openai- prefix).
+  //
+  // Compaction is a structured transcription job, not a reasoning task — the
+  // P7 prompt's three passes (USER VOICE / EVENT LOG / STATE NARRATIVE)
+  // produce ~1.5K output tokens of well-shaped text. Using reasoning models
+  // for this burns reasoning-token cost on a task that doesn't need it AND
+  // adds significant TTFT/total latency per pass. The previous defaults
+  // (gpt-5.1-codex-mini, gpt-5-mini, gpt-5-nano, o4-mini — all reasoning)
+  // were chosen for their 400K context windows, but the gpt-4.1 family
+  // offers 1M-token context AND is non-reasoning AND ~30× cheaper than the
+  // previous primary.
+  //
+  // Both entries below have 1M context, so the runtime model-filter never
+  // has to fall through to a smaller-context model for size reasons.
   COMPACTION_MODELS: [
-    'gpt-5.1-codex-mini',           // 400K context - best for large conversations
-    'gpt-5-mini',                   // 400K context
-    'gpt-5-nano',                   // 400K context - lightweight
-    'o4-mini'                       // 128K context - reasoning model
+    'gpt-4.1-nano',                 // 1M context, non-reasoning, cheapest — primary
+    'gpt-4.1-mini',                 // 1M context, non-reasoning           — fallback
   ],
 
   // Context windows for recommended compaction models (fallback if modelsService unavailable)
   MODEL_CONTEXT_WINDOWS: {
-    'gpt-5.1-codex-mini': 400000,
-    'gpt-5-mini': 400000,
-    'gpt-5-nano': 400000,
-    'o4-mini': 128000
+    'gpt-4.1-nano': 1048576,
+    'gpt-4.1-mini': 1048576,
   },
 
   // Token limits

package/web-ui/build/index.html

@@ -6,8 +6,8 @@
     <meta name="viewport" content="width=device-width, initial-scale=1.0" />
     <title>Loxia Autopilot One</title>
     <meta name="description" content="AI Agents System - No-code/vibe-code/companion-coder platform" />
-    <script type="module" crossorigin src="/static/index-whZPU4as.js"></script>
-    <link rel="stylesheet" crossorigin href="/static/index-DEWSWosh.css">
+    <script type="module" crossorigin src="/static/index-BBHrMvg7.js"></script>
+    <link rel="stylesheet" crossorigin href="/static/index-_z8JjNrI.css">
   </head>
   <body>
     <div id="root"></div>

package/web-ui/build/static/1c-BoCdC5pC.js

@@ -0,0 +1 @@
+import{g as f}from"./index-BBHrMvg7.js";import{r as s}from"./index-ChkZR_aE.js";function i(t,n){for(var o=0;o<n.length;o++){const e=n[o];if(typeof e!="string"&&!Array.isArray(e)){for(const r in e)if(r!=="default"&&!(r in t)){const c=Object.getOwnPropertyDescriptor(e,r);c&&Object.defineProperty(t,r,c.get?c:{enumerable:!0,get:()=>e[r]})}}}return Object.freeze(Object.defineProperty(t,Symbol.toStringTag,{value:"Module"}))}var a=s();const p=f(a),l=i({__proto__:null,default:p},[a]);export{l as _};

package/web-ui/build/static/abap-D7A92M1K.js

@@ -0,0 +1 @@
+import{g as s,r as f}from"./index-BBHrMvg7.js";function c(t,n){for(var o=0;o<n.length;o++){const e=n[o];if(typeof e!="string"&&!Array.isArray(e)){for(const r in e)if(r!=="default"&&!(r in t)){const a=Object.getOwnPropertyDescriptor(e,r);a&&Object.defineProperty(t,r,a.get?a:{enumerable:!0,get:()=>e[r]})}}}return Object.freeze(Object.defineProperty(t,Symbol.toStringTag,{value:"Module"}))}var p=f();const i=s(p),b=c({__proto__:null,default:i},[p]);export{b as a};

package/web-ui/build/static/abnf-DIcpJ99N.js

@@ -0,0 +1 @@
+import{g as s,d as c}from"./index-BBHrMvg7.js";function i(t,a){for(var o=0;o<a.length;o++){const e=a[o];if(typeof e!="string"&&!Array.isArray(e)){for(const r in e)if(r!=="default"&&!(r in t)){const n=Object.getOwnPropertyDescriptor(e,r);n&&Object.defineProperty(t,r,n.get?n:{enumerable:!0,get:()=>e[r]})}}}return Object.freeze(Object.defineProperty(t,Symbol.toStringTag,{value:"Module"}))}var f=c();const p=s(f),b=i({__proto__:null,default:p},[f]);export{b as a};

package/web-ui/build/static/abnf-d4I9yABN.js

@@ -0,0 +1 @@
+import{g as s}from"./index-BBHrMvg7.js";import{a as i}from"./index-ChkZR_aE.js";function c(t,n){for(var o=0;o<n.length;o++){const e=n[o];if(typeof e!="string"&&!Array.isArray(e)){for(const r in e)if(r!=="default"&&!(r in t)){const a=Object.getOwnPropertyDescriptor(e,r);a&&Object.defineProperty(t,r,a.get?a:{enumerable:!0,get:()=>e[r]})}}}return Object.freeze(Object.defineProperty(t,Symbol.toStringTag,{value:"Module"}))}var f=i();const p=s(f),g=c({__proto__:null,default:p},[f]);export{g as a};

package/web-ui/build/static/accesslog-D_Ksv5Fi.js

@@ -0,0 +1 @@
+import{g as n}from"./index-BBHrMvg7.js";import{b as f}from"./index-ChkZR_aE.js";function g(t,c){for(var o=0;o<c.length;o++){const e=c[o];if(typeof e!="string"&&!Array.isArray(e)){for(const r in e)if(r!=="default"&&!(r in t)){const s=Object.getOwnPropertyDescriptor(e,r);s&&Object.defineProperty(t,r,s.get?s:{enumerable:!0,get:()=>e[r]})}}}return Object.freeze(Object.defineProperty(t,Symbol.toStringTag,{value:"Module"}))}var a=f();const l=n(a),u=g({__proto__:null,default:l},[a]);export{u as a};

package/web-ui/build/static/actionscript-CMJI1an7.js

@@ -0,0 +1 @@
+import{g as a}from"./index-BBHrMvg7.js";import{c as s}from"./index-ChkZR_aE.js";function p(r,i){for(var o=0;o<i.length;o++){const t=i[o];if(typeof t!="string"&&!Array.isArray(t)){for(const e in t)if(e!=="default"&&!(e in r)){const c=Object.getOwnPropertyDescriptor(t,e);c&&Object.defineProperty(r,e,c.get?c:{enumerable:!0,get:()=>t[e]})}}}return Object.freeze(Object.defineProperty(r,Symbol.toStringTag,{value:"Module"}))}var n=s();const f=a(n),l=p({__proto__:null,default:f},[n]);export{l as a};

package/web-ui/build/static/actionscript-CNcWu4a1.js

@@ -0,0 +1 @@
+import{g as a,e as s}from"./index-BBHrMvg7.js";function p(r,i){for(var o=0;o<i.length;o++){const t=i[o];if(typeof t!="string"&&!Array.isArray(t)){for(const e in t)if(e!=="default"&&!(e in r)){const c=Object.getOwnPropertyDescriptor(t,e);c&&Object.defineProperty(r,e,c.get?c:{enumerable:!0,get:()=>t[e]})}}}return Object.freeze(Object.defineProperty(r,Symbol.toStringTag,{value:"Module"}))}var n=s();const f=a(n),g=p({__proto__:null,default:f},[n]);export{g as a};

package/web-ui/build/static/ada-BIqOBKro.js

@@ -0,0 +1 @@
+import{g as s,f as c}from"./index-BBHrMvg7.js";function i(t,n){for(var o=0;o<n.length;o++){const e=n[o];if(typeof e!="string"&&!Array.isArray(e)){for(const r in e)if(r!=="default"&&!(r in t)){const a=Object.getOwnPropertyDescriptor(e,r);a&&Object.defineProperty(t,r,a.get?a:{enumerable:!0,get:()=>e[r]})}}}return Object.freeze(Object.defineProperty(t,Symbol.toStringTag,{value:"Module"}))}var f=c();const p=s(f),d=i({__proto__:null,default:p},[f]);export{d as a};

package/web-ui/build/static/ada-BaTkJMdz.js

@@ -0,0 +1 @@
+import{g as s}from"./index-BBHrMvg7.js";import{d as i}from"./index-ChkZR_aE.js";function c(t,n){for(var o=0;o<n.length;o++){const e=n[o];if(typeof e!="string"&&!Array.isArray(e)){for(const r in e)if(r!=="default"&&!(r in t)){const a=Object.getOwnPropertyDescriptor(e,r);a&&Object.defineProperty(t,r,a.get?a:{enumerable:!0,get:()=>e[r]})}}}return Object.freeze(Object.defineProperty(t,Symbol.toStringTag,{value:"Module"}))}var f=i();const p=s(f),g=c({__proto__:null,default:p},[f]);export{g as a};

package/web-ui/build/static/agda-DmRdGgvt.js

@@ -0,0 +1 @@
+import{g as f,h as g}from"./index-BBHrMvg7.js";function c(t,n){for(var o=0;o<n.length;o++){const e=n[o];if(typeof e!="string"&&!Array.isArray(e)){for(const r in e)if(r!=="default"&&!(r in t)){const a=Object.getOwnPropertyDescriptor(e,r);a&&Object.defineProperty(t,r,a.get?a:{enumerable:!0,get:()=>e[r]})}}}return Object.freeze(Object.defineProperty(t,Symbol.toStringTag,{value:"Module"}))}var s=g();const i=f(s),u=c({__proto__:null,default:i},[s]);export{u as a};