v1.13.3: cleanup bundle — statement timeout + alpha ordering + stuck-row sweeper + repairToolCall

Four independent items, all owed from prior dispatches.

- statement_timeout at the database level via:
    ALTER DATABASE boocode SET statement_timeout = '30s';
  Applied operationally; documented as a comment at the top of schema.sql
  (ALTER DATABASE can't run inside a DO block, so it's not idempotent
  inside applySchema). Re-apply after a volume reset.

- Tool registry alpha-sorted at module load. llama.cpp's prompt cache
  hits on byte-identical prefixes; any reordering of the tool list near
  the top of the system prompt would invalidate every cached turn.
  Single-source sort at the ALL_TOOLS export so toolJsonSchemas() and
  TOOLS_BY_NAME inherit the order automatically. New tools.test.ts
  asserts the invariant; total tests 173 (was 172).

- Periodic in-process stuck-row sweeper. Runs every 60s, marks
  'streaming' rows older than 5 minutes as 'failed', and publishes
  chat_status='idle' on the user channel so the UI dot drops without a
  refresh. Closes the mid-session crash UX gap; the v1.12.1 boot sweep
  only fires once at startup, so sessions used to stay stuck until next
  reboot. setInterval cleaned up via app.addHook('onClose'). Mirrors
  handleAbortOrError's publish pattern.

- experimental_repairToolCall wired through AI SDK v6 streamText. Pass-
  through implementation: log + return the original toolCall so the
  stream keeps going. executeToolPhase's existing error paths (unknown
  tool name → 'unknown tool: X' result; zod-reject → 'tool X rejected
  — field: required') already surface bad calls to the model; the value
  here is preventing the AI SDK from THROWING on parse errors and
  killing the whole stream. Owed since v1.13.1-A.

Smoke verified:
- statement_timeout = '30s' confirmed via SHOW.
- Tool path normal flow intact (list_dir prompt → tool_call → result
  → final assistant). No malformed tool calls in the test run; repair
  log will surface them when qwen3.6 actually emits one.
- Alpha order verified at runtime via the dist bundle: match: true.
- Sweeper logic not traffic-tested (no stuck rows to find), but the
  SQL UPDATE + broker.publishUser pattern is identical to handleAbort
  and the boot sweep — synthesis-only verification.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

CLAUDE.md

@@ -46,7 +46,9 @@ Tests: `pnpm -C apps/server test` runs the vitest suite. No test harness on `app
 - **Zod** for request validation and config parsing.
 
 Key services:
-- **`services/inference.ts`** — Streams LLM responses, executes tool loops (max depth 15, see `MAX_TOOL_LOOP_DEPTH`), flushes to DB every 500ms. Publishes `InferenceFrame` events through the broker. **`TurnArgs`** is the per-turn state envelope threaded through the `executeToolPhase → runAssistantTurn` recursion (`toolsUsed`, `recentToolCalls`, `assistantMessageId`, `signal`); reset to defaults in `runInference` at the user-message boundary. Cap-hit (`toolsUsed >= budget`) and doom-loop (`detectDoomLoop(recentToolCalls)`) checks both read from this envelope. Add new per-turn state here, not in module-level closures.
+- **`services/inference/`** (v1.12.4 split — was a single `inference.ts` file). Public surface re-exported via `inference/index.ts`; callers import from `./services/inference/index.js`. Layout: `turn.ts` (runAssistantTurn / runInference / createInferenceRunner orchestration, plus `InferenceFrame`, `InferenceContext`, `TurnArgs`, `StreamResult` exported), `stream-phase.ts` (streamCompletion + executeStreamPhase + SSE parsing), `tool-phase.ts` (executeToolPhase; back-edges into turn.ts for the runAssistantTurn recursion — cycle is safe because dereferenced at call time, not module top-level), `sentinel-summaries.ts` (runCapHitSummary + runDoomLoopSummary + their sentinel inserters; two near-clones kept side-by-side until a third sentinel justifies factoring out runWrapUpSummary), `error-handler.ts` (handleAbortOrError, finalizeCompletion), `payload.ts` (buildMessagesPayload, loadContext, maybeFlagForCompaction, `OpenAiMessage`), `sentinels.ts` (`detectDoomLoop`, `DOOM_LOOP_THRESHOLD`, sentinel predicates), `budget.ts` (resolveToolBudget), `xml-parser.ts` (Qwen-coder XML tool-call fallback), `types.ts` (`StreamPhaseState`, `DB_FLUSH_INTERVAL_MS` shared between stream-phase and sentinel-summaries). **`TurnArgs`** is the per-turn state envelope threaded through the `executeToolPhase → runAssistantTurn` recursion (`toolsUsed`, `recentToolCalls`, `assistantMessageId`, `signal`); reset to defaults in `runInference` at the user-message boundary. Cap-hit (`toolsUsed >= budget`) and doom-loop (`detectDoomLoop(recentToolCalls)`) checks both read from this envelope. Add new per-turn state to `TurnArgs` in `turn.ts`, not module-level closures.
+- **`chat_status` frame shape** (published via `broker.publishUser`) — `status: 'streaming' | 'tool_running' | 'waiting_for_input' | 'idle' | 'error'` (widened from `working|idle|error` in v1.12.1). Frontend `useChatStatus` derives `idle_warm` (<30s since idle) vs `idle_cold`. `ChatThroughput` renders inline beside `StatusDot` only when streaming or tool_running, fed by 500ms-throttled `'usage'` WS frames (`completion_tokens` + `ctx_used` + `ctx_max`). The `POST /api/chats/:id/discard_stale` endpoint exists to mark a stuck-streaming row as `failed` when the frontend's 60s no-token-activity timer (`ChatPane` content-length watcher) gives up.
+- **Boot-time stale-streaming sweep** in `apps/server/src/index.ts` after `applySchema()`: any `messages.status='streaming'` older than 5 minutes flips to `'failed'`. Logs only on non-zero count. Recovers from container restart while inference was mid-stream (v1.12.1).
 - **`services/broker.ts`** — In-memory pub/sub with two channel types: per-session (message streaming) and per-user (sidebar updates). No persistence; clients reconnect on restart.
 - **`services/tools.ts`** — Tool registry (`ALL_TOOLS`, `READ_ONLY_TOOL_NAMES`, `TOOLS_BY_NAME`). Filesystem tools (view_file/list_dir/grep/find_files) go through three guard layers: `path_guard.ts` (workspace scope), `secret_guard.ts` (filename deny list), `url_guard.ts` (SSRF/private-IP block for web_fetch). v1.11.8+ web tools (`web_search`, `web_fetch`) are opt-in per chat via `session.web_search_enabled` (resolved with `project.default_web_search_enabled` fallback) and filtered out of the LLM's tool schema when false.
 - **`services/compaction.ts`** + **`services/model-context.ts`** — v1.11.0 anchored rolling summary (single `summary=true` assistant row per chat, supersedes itself on each compaction). Triggered when `chats.needs_compaction` is set after an inference turn exceeds `usable(ctx_max) = ctx_max - 20k`. **`ctx_max` comes from `model-context.getModelContext()` which fetches `${LLAMA_SWAP_URL}/upstream/<model>/props`** — NOT from `parsed.timings.n_ctx` (the stream completion's `timings` doesn't carry n_ctx; that read was dead code until v1.11.3 ripped it out).
@@ -87,15 +89,14 @@ Font / CSS pipeline (apps/web):
 
 ### Multi-pane workspace
 
-Sessions hold 1–5 panes (chat / empty / placeholder terminal+agent). Workspace pane state is **client-side only** (localStorage key `boocode.workspace.panes.<sessionId>`); the legacy `session_panes` table and its REST endpoints are deprecated — no `/api/panes/*` routes exist. Each chat lives in at most one pane; tab strip is per-pane and tracks `chatIds[]` + `activeChatIdx`. Sessions 1:N chats; chats own messages. Tab reorder via native HTML5 drag events.
+Sessions hold 1–5 panes (chat / empty / placeholder terminal+agent). v1.12.1 moved pane state from per-device localStorage to `sessions.workspace_panes jsonb` for cross-device sync. `PATCH /api/sessions/:id/workspace` persists; `session_workspace_updated` user-channel frame broadcasts to every device watching the session. `useWorkspacePanes` debounces saves 300ms and dedups echoes by JSON string. Legacy localStorage key `boocode.workspace.panes.<sessionId>` is read once on first hydrate (one-time seed-and-delete migration when server is empty but localStorage has data); no longer written. The deprecated `session_panes` table was dropped. `validatePanes(validChatIds)` prunes panes referencing chat IDs that no longer exist (called by `useSessionChats` after the chat list fetch lands). Each chat lives in at most one pane; tab strip is per-pane and tracks `chatIds[]` + `activeChatIdx`. Tab reorder via native HTML5 drag events.
 
 ## Database
 
-PostgreSQL 16. Tables: `projects`, `sessions`, `chats`, `messages`, `settings`, `session_panes` (deprecated). Schema applied idempotently on startup via `applySchema()`. Use `clock_timestamp()` (not `NOW()`) inside transactions. CHECK constraints in place: `projects_status_chk` ('open'|'archived'), `sessions_status_chk` (same), `chats_status_chk` (same), `messages_role_chk`, `messages_status_chk` — keep in sync with the `*_STATUSES` const arrays in `apps/server/src/types/api.ts`.
+PostgreSQL 16. Tables: `projects`, `sessions`, `chats`, `messages`, `settings`. (`session_panes` was dropped in v1.12.1; workspace pane state lives in `sessions.workspace_panes jsonb`.) Schema applied idempotently on startup via `applySchema()`. Use `clock_timestamp()` (not `NOW()`) inside transactions. CHECK constraints in place: `projects_status_chk` ('open'|'archived'), `sessions_status_chk` (same), `chats_status_chk` (same), `messages_role_chk`, `messages_status_chk` — keep in sync with the `*_STATUSES` const arrays in `apps/server/src/types/api.ts`. The older anonymous `messages_status_check` (without 'cancelled') and `messages_role_check` (without 'system') were dropped in v1.12.1; only the `_chk` variants remain.
 
 Schema CHECK migration order when renaming allowed values: (1) `ALTER TABLE ... DROP CONSTRAINT IF EXISTS <system_name>` (inline `CREATE TABLE` checks get `<table>_<column>_check`), (2) `UPDATE` rows to new values, (3) wrap new constraint ADD in `DO $$ ... pg_constraint` guard — that block is the only way to get `ADD CONSTRAINT IF NOT EXISTS`.
 
-Position-shift pattern for panes (legacy `session_panes` table): negate-and-restore to avoid UNIQUE(session_id, position) collisions during reorder/insert/delete. Sentinel value -100 for the moving pane.
 
 ## Environment
 
@@ -125,6 +126,7 @@ Required: `DATABASE_URL`, `LLAMA_SWAP_URL`. Optional: `PORT` (3000), `HOST` (0.0
 - TypeScript strict mode. Both apps share `tsconfig.base.json`.
 - Server uses NodeNext module resolution (`.js` extensions in imports).
 - Discriminated unions for type narrowing: `Pane` (by `kind`), `SessionEvent` (by `type`), `InferenceFrame` (by `type`).
+- **Adding a new WS frame type** requires updating BOTH the server's `InferenceFrame` (loose `type:` union + optional fields in `services/inference/turn.ts`) AND the web `WsFrame` (strict discriminated union in `apps/web/src/api/types.ts`). Server publish is permissive; the frontend type is the wire-format gate. The `'usage'` frame added in v1.12.2 needed both sides; missing the web side silently drops the frame at JSON-parse.
 - shadcn primitives live in `components/ui/`. Don't modify them unless adding a new primitive.
 - `inferLanguage()` from `lib/attachments.ts` is the canonical file-extension-to-language map. `CodeBlock.tsx` keeps its own `LANG_MAP` because it also resolves markdown fence names.
 - Two UI event buses: `hooks/sessionEvents.ts` for DB-state events (chat_created, session_updated); `lib/events.ts` for ephemeral UI (`sendToTerminal`, `terminalsRegistry`). Don't merge — different subscriber lifecycles.

apps/server/package.json

@@ -11,8 +11,10 @@
     "test": "vitest run"
   },
   "dependencies": {
+    "@ai-sdk/openai-compatible": "^2.0.47",
     "@fastify/static": "^7.0.4",
     "@fastify/websocket": "^10.0.1",
+    "ai": "^6.0.190",
     "fastify": "^4.28.1",
     "postgres": "^3.4.4",
     "ws": "^8.18.0",

apps/server/src/index.ts

@@ -16,7 +16,7 @@ import { registerWebSocket } from './routes/ws.js';
 import { registerModelRoutes } from './routes/models.js';
 import { registerAgentRoutes } from './routes/agents.js';
 import { registerSkillsRoutes } from './routes/skills.js';
-import { createInferenceRunner } from './services/inference.js';
+import { createInferenceRunner } from './services/inference/index.js';
 import { createBroker } from './services/broker.js';
 import { listSkills } from './services/skills.js';
 import * as compaction from './services/compaction.js';
@@ -201,6 +201,46 @@ async function main() {
     app.log.info(`serving static frontend from ${webDist}`);
   }
 
+  // v1.13.3: periodic in-process sweeper for streaming rows orphaned by a
+  // mid-session crash. The boot sweep (above) only fires once at startup;
+  // this loop catches the in-flight case. 60s cadence + 5-min threshold
+  // matches the boot sweep so behavior is consistent. Publishes
+  // chat_status='idle' on the user channel so the UI dot drops without a
+  // refresh — same pattern as handleAbortOrError.
+  const SWEEP_INTERVAL_MS = 60_000;
+  const sweepStaleStreaming = async (): Promise<void> => {
+    try {
+      const rows = await sql<{ id: string; chat_id: string }[]>`
+        UPDATE messages
+        SET status = 'failed', finished_at = clock_timestamp()
+        WHERE status = 'streaming'
+          AND created_at < NOW() - INTERVAL '5 minutes'
+        RETURNING id, chat_id
+      `;
+      if (rows.length === 0) return;
+      app.log.warn(
+        { swept: rows.length, ids: rows.map((r) => r.id) },
+        'swept stale streaming rows',
+      );
+      const seenChats = new Set<string>();
+      const now = new Date().toISOString();
+      for (const row of rows) {
+        if (seenChats.has(row.chat_id)) continue;
+        seenChats.add(row.chat_id);
+        broker.publishUser('default', {
+          type: 'chat_status',
+          chat_id: row.chat_id,
+          status: 'idle',
+          at: now,
+        });
+      }
+    } catch (err) {
+      app.log.error({ err }, 'stuck-row sweeper failed');
+    }
+  };
+  const sweepTimer = setInterval(() => { void sweepStaleStreaming(); }, SWEEP_INTERVAL_MS);
+  app.addHook('onClose', async () => { clearInterval(sweepTimer); });
+
   const shutdown = async (signal: string) => {
     app.log.info(`received ${signal}, shutting down`);
     try {

apps/server/src/routes/chats.ts

@@ -18,6 +18,12 @@ const ForkBody = z.object({
   name: z.string().min(1).max(200).optional(),
 });
 
+const DiscardStaleBody = z.object({
+  message_id: z.string().uuid(),
+});
+
+const STALE_MIN_AGE_SECONDS = 60;
+
 export function registerChatRoutes(
   app: FastifyInstance,
   sql: Sql,
@@ -307,6 +313,28 @@ export function registerChatRoutes(
             AND created_at <= ${target.created_at}::timestamptz
             AND status = 'complete'
         `;
+        // v1.13.0: clone message_parts for the forked messages. Source and
+        // destination preserve ordering (the INSERT above orders by created_at,
+        // id) so a ROW_NUMBER pairing maps source.id → dest.id deterministically.
+        await tx`
+          WITH src AS (
+            SELECT id, ROW_NUMBER() OVER (ORDER BY created_at ASC, id ASC) AS rn
+            FROM messages
+            WHERE chat_id = ${source.id}
+              AND created_at <= ${target.created_at}::timestamptz
+              AND status = 'complete'
+          ),
+          dst AS (
+            SELECT id, ROW_NUMBER() OVER (ORDER BY created_at ASC, id ASC) AS rn
+            FROM messages
+            WHERE chat_id = ${chat!.id}
+          )
+          INSERT INTO message_parts (message_id, sequence, kind, payload)
+          SELECT dst.id, p.sequence, p.kind, p.payload
+          FROM message_parts p
+          JOIN src ON p.message_id = src.id
+          JOIN dst ON dst.rn = src.rn
+        `;
         return chat!;
       });
 
@@ -320,6 +348,73 @@ export function registerChatRoutes(
     }
   );
 
+  // v1.12.3: explicit recovery from a stuck-streaming assistant row. The
+  // frontend gates this behind a 60s no-token-activity timer; the server
+  // re-checks the age and current status for safety. Non-streaming rows
+  // return 409 (frontend race; idempotent retry is fine).
+  app.post<{ Params: { id: string } }>(
+    '/api/chats/:id/discard_stale',
+    async (req, reply) => {
+      const parsed = DiscardStaleBody.safeParse(req.body ?? {});
+      if (!parsed.success) {
+        reply.code(400);
+        return { error: 'invalid body', details: parsed.error.flatten() };
+      }
+      const rows = await sql<{
+        id: string;
+        session_id: string;
+        chat_id: string;
+        status: string;
+        age_seconds: number;
+      }[]>`
+        SELECT id, session_id, chat_id, status,
+               EXTRACT(EPOCH FROM (clock_timestamp() - created_at))::int AS age_seconds
+        FROM messages
+        WHERE id = ${parsed.data.message_id} AND chat_id = ${req.params.id}
+      `;
+      if (rows.length === 0) {
+        reply.code(404);
+        return { error: 'message not found in chat' };
+      }
+      const msg = rows[0]!;
+      if (msg.status !== 'streaming') {
+        reply.code(409);
+        return { error: 'message is no longer streaming', current_status: msg.status };
+      }
+      if (msg.age_seconds < STALE_MIN_AGE_SECONDS) {
+        reply.code(409);
+        return { error: 'message is not stale yet', age_seconds: msg.age_seconds };
+      }
+      const updated = await sql<Message[]>`
+        UPDATE messages
+        SET status = 'failed',
+            content = COALESCE(content, ''),
+            finished_at = clock_timestamp()
+        WHERE id = ${msg.id} AND status = 'streaming'
+        RETURNING id, session_id, chat_id, role, content, kind, tool_calls, tool_results,
+                  status, last_seq, tokens_used, ctx_used, ctx_max, started_at, finished_at,
+                  created_at, metadata, summary, tail_start_id, compacted_at
+      `;
+      if (updated.length === 0) {
+        // Race: the row flipped out of 'streaming' between our SELECT and UPDATE.
+        reply.code(409);
+        return { error: 'message status changed mid-request' };
+      }
+      broker.publishUser('default', {
+        type: 'chat_status',
+        chat_id: msg.chat_id,
+        status: 'idle',
+        at: new Date().toISOString(),
+      });
+      broker.publish(msg.session_id, {
+        type: 'message_complete',
+        message_id: msg.id,
+        chat_id: msg.chat_id,
+      });
+      return updated[0];
+    }
+  );
+
   app.get<{ Params: { id: string } }>(
     '/api/chats/:id/messages',
     async (req, reply) => {
@@ -328,11 +423,12 @@ export function registerChatRoutes(
         reply.code(404);
         return { error: 'chat not found' };
       }
+      // v1.13.1-B: reads tool_calls/tool_results via the parts-merged view.
       const rows = await sql<Message[]>`
         SELECT id, session_id, chat_id, role, content, kind, tool_calls, tool_results, status, last_seq,
                tokens_used, ctx_used, ctx_max, started_at, finished_at, created_at, metadata,
                summary, tail_start_id, compacted_at
-        FROM messages
+        FROM messages_with_parts
         WHERE chat_id = ${req.params.id}
         ORDER BY created_at ASC, id ASC
       `;

apps/server/src/routes/messages.ts

@@ -91,11 +91,12 @@ export function registerMessageRoutes(
       // SummaryCard) and shows compacted_at-stamped rows inline for context.
       // Internal inference assembly filters compacted_at IS NULL separately —
       // see services/inference.ts loadContext + services/compaction.ts.
+      // v1.13.1-B: reads tool_calls/tool_results via the parts-merged view.
       const rows = await sql<Message[]>`
         SELECT id, session_id, chat_id, role, content, kind, tool_calls, tool_results, status, last_seq,
                tokens_used, ctx_used, ctx_max, started_at, finished_at, created_at, metadata,
                summary, tail_start_id, compacted_at
-        FROM messages
+        FROM messages_with_parts
         WHERE session_id = ${req.params.id}
         ORDER BY created_at ASC, id ASC
       `;
@@ -469,30 +470,36 @@ export function registerMessageRoutes(
       const chat = chatRows[0]!;
       const sessionId = chat.session_id;
 
-      // Find the assistant message that emitted this tool_call. Scoped by
-      // chat_id + role to avoid cross-chat lookups; ordered by created_at DESC
-      // because the most recent issuance wins when an LLM reuses call IDs
-      // across turns (the older, already-answered one is a different row with
-      // populated tool_results downstream).
-      const callerRows = await sql<{ id: string; tool_calls: ToolCall[] | null }[]>`
-        SELECT id, tool_calls FROM messages
-        WHERE chat_id = ${chat.id}
-          AND role = 'assistant'
-          AND tool_calls IS NOT NULL
-        ORDER BY created_at DESC
+      // v1.13.1-C: find the assistant's tool_call by indexing message_parts
+      // directly on payload->>'id'. Scoped by chat_id + role via the JOIN.
+      // Pre-v1.13.0 history has no parts rows — those tool_calls become
+      // unreachable here (404). Acceptable per the dispatch decision: any
+      // pending elicitation from before v1.13.0 is long timed out by now;
+      // promote to a hotfix with a JSON-column fallback if it ever surfaces.
+      const callerRows = await sql<{
+        message_id: string;
+        payload: { id: string; name: string; args: Record<string, unknown> };
+      }[]>`
+        SELECT p.message_id, p.payload
+        FROM message_parts p
+        JOIN messages m ON m.id = p.message_id
+        WHERE m.chat_id = ${chat.id}
+          AND m.role = 'assistant'
+          AND p.kind = 'tool_call'
+          AND p.payload->>'id' = ${tool_call_id}
+        ORDER BY m.created_at DESC
+        LIMIT 1
       `;
-      let foundCall: ToolCall | null = null;
-      for (const row of callerRows) {
-        const match = row.tool_calls?.find((tc) => tc.id === tool_call_id);
-        if (match) {
-          foundCall = match;
-          break;
-        }
-      }
-      if (!foundCall) {
+      const callerRow = callerRows[0];
+      if (!callerRow) {
         reply.code(404);
         return { error: 'unknown_tool_call_id' };
       }
+      const foundCall: ToolCall = {
+        id: callerRow.payload.id,
+        name: callerRow.payload.name,
+        args: callerRow.payload.args,
+      };
       if (foundCall.name !== 'ask_user_input') {
         reply.code(400);
         return { error: 'tool_call_not_ask_user_input' };
@@ -539,18 +546,21 @@ export function registerMessageRoutes(
         }
       }
 
-      // Find the pending tool row. ORDER BY created_at DESC + LIMIT 1 picks
-      // the most recent row with this tool_call_id; the already-answered
-      // check below guards against UPDATE-ing a stale answer.
+      // v1.13.1-C: find the pending tool row via message_parts on
+      // payload->>'tool_call_id'. Same fallback caveat as the caller lookup
+      // above — pre-v1.13.0 rows are unreachable here.
       const toolRows = await sql<{
-        id: string;
-        tool_results: { tool_call_id: string; output: unknown } | null;
+        message_id: string;
+        payload: { tool_call_id: string; output: unknown };
       }[]>`
-        SELECT id, tool_results FROM messages
-        WHERE chat_id = ${chat.id}
-          AND role = 'tool'
-          AND tool_results->>'tool_call_id' = ${tool_call_id}
-        ORDER BY created_at DESC
+        SELECT p.message_id, p.payload
+        FROM message_parts p
+        JOIN messages m ON m.id = p.message_id
+        WHERE m.chat_id = ${chat.id}
+          AND m.role = 'tool'
+          AND p.kind = 'tool_result'
+          AND p.payload->>'tool_call_id' = ${tool_call_id}
+        ORDER BY m.created_at DESC
         LIMIT 1
       `;
       const toolRow = toolRows[0];
@@ -558,7 +568,7 @@ export function registerMessageRoutes(
         reply.code(404);
         return { error: 'unknown_tool_call_id', detail: 'tool message not found' };
       }
-      if (toolRow.tool_results && toolRow.tool_results.output !== null) {
+      if (toolRow.payload && toolRow.payload.output !== null) {
         reply.code(409);
         return { error: 'tool_call_already_answered' };
       }
@@ -570,11 +580,21 @@ export function registerMessageRoutes(
         truncated: false,
       };
 
+      const toolMessageId = toolRow.message_id;
       const result = await sql.begin(async (tx) => {
         await tx`
           UPDATE messages
           SET tool_results = ${tx.json(newToolResults as never)}
-          WHERE id = ${toolRow.id}
+          WHERE id = ${toolMessageId}
+        `;
+        // v1.13.0: replace the pending tool_result part inserted at message
+        // creation (tool-phase.ts) with the answered one. Delete-then-insert
+        // is simpler than UPDATE because parts are append-style elsewhere;
+        // the UNIQUE (message_id, sequence) constraint blocks plain insert.
+        await tx`DELETE FROM message_parts WHERE message_id = ${toolMessageId} AND kind = 'tool_result'`;
+        await tx`
+          INSERT INTO message_parts (message_id, sequence, kind, payload)
+          VALUES (${toolMessageId}, 0, 'tool_result', ${tx.json(newToolResults as never)})
         `;
         const [assistantMsg] = await tx<{ id: string }[]>`
           INSERT INTO messages (session_id, chat_id, role, content, status, created_at)
@@ -584,7 +604,7 @@ export function registerMessageRoutes(
         await tx`UPDATE sessions SET updated_at = clock_timestamp() WHERE id = ${sessionId}`;
         await tx`UPDATE chats SET updated_at = clock_timestamp() WHERE id = ${chat.id}`;
         return {
-          tool_message_id: toolRow.id,
+          tool_message_id: toolMessageId,
           assistant_message_id: assistantMsg!.id,
         };
       });

apps/server/src/routes/skills.ts

@@ -90,11 +90,26 @@ export function registerSkillsRoutes(
           VALUES (${sessionId}, ${chat.id}, 'assistant', '', ${sql.json(toolCalls as never)}, 'complete', clock_timestamp())
           RETURNING id
         `;
+        // v1.13.0: dual-write the synthetic assistant message's tool_call.
+        // Single skill_use tool_call, no text content, so one part at seq 0.
+        await tx`
+          INSERT INTO message_parts (message_id, sequence, kind, payload)
+          VALUES (${synthAssistant!.id}, 0, 'tool_call', ${tx.json({
+            id: toolCallId,
+            name: 'skill_use',
+            args: { name: skill_name },
+          } as never)})
+        `;
         const [toolMsg] = await tx<{ id: string }[]>`
           INSERT INTO messages (session_id, chat_id, role, content, tool_results, status, created_at)
           VALUES (${sessionId}, ${chat.id}, 'tool', '', ${sql.json(toolResults as never)}, 'complete', clock_timestamp())
           RETURNING id
         `;
+        // v1.13.0: dual-write the synthetic tool result (the skill body).
+        await tx`
+          INSERT INTO message_parts (message_id, sequence, kind, payload)
+          VALUES (${toolMsg!.id}, 0, 'tool_result', ${tx.json(toolResults as never)})
+        `;
         const [userMsg] = await tx<{ id: string }[]>`
           INSERT INTO messages (session_id, chat_id, role, content, status, created_at)
           VALUES (${sessionId}, ${chat.id}, 'user', ${userText}, 'complete', clock_timestamp())

apps/server/src/routes/ws.ts

@@ -23,11 +23,12 @@ export function registerWebSocket(
 
       // v1.11: snapshot includes compaction fields so MessageBubble can
       // render the SummaryCard for summary=true rows on first connect.
+      // v1.13.1-B: reads tool_calls/tool_results via the parts-merged view.
       const messages = await sql<Message[]>`
         SELECT id, session_id, chat_id, role, content, kind, tool_calls, tool_results, status, last_seq,
                tokens_used, ctx_used, ctx_max, started_at, finished_at, created_at, metadata,
                summary, tail_start_id, compacted_at
-        FROM messages
+        FROM messages_with_parts
         WHERE session_id = ${sessionId}
         ORDER BY created_at ASC, id ASC
       `;

apps/server/src/schema.sql

@@ -1,3 +1,10 @@
+-- v1.13.3: statement_timeout is set at database level via:
+--   ALTER DATABASE boocode SET statement_timeout = '30s';
+-- ALTER DATABASE can't run inside a DO block, so this is an operational
+-- step rather than schema. Re-apply after a volume reset (the setting
+-- lives in pg_db which survives `docker compose up --build` but NOT a
+-- `docker volume rm boocode_pgdata`).
+
 CREATE TABLE IF NOT EXISTS projects (
   id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
   name TEXT NOT NULL,
@@ -32,6 +39,59 @@ CREATE TABLE IF NOT EXISTS messages (
 
 CREATE INDEX IF NOT EXISTS idx_messages_session ON messages(session_id, created_at);
 
+-- v1.13.0: granular message parts table for AI SDK migration. Old
+-- messages.content / tool_calls / tool_results columns stay authoritative
+-- for reads in v1.13.0; this table is dual-written so the swap can happen
+-- in a later dispatch without a backfill window. ON DELETE CASCADE means
+-- removing a message removes its parts in one go.
+CREATE TABLE IF NOT EXISTS message_parts (
+  id uuid PRIMARY KEY DEFAULT gen_random_uuid(),
+  message_id uuid NOT NULL REFERENCES messages(id) ON DELETE CASCADE,
+  sequence int NOT NULL,
+  kind text NOT NULL,
+  payload jsonb NOT NULL,
+  created_at timestamptz NOT NULL DEFAULT clock_timestamp(),
+  CONSTRAINT message_parts_kind_chk CHECK (kind IN ('text', 'tool_call', 'tool_result', 'reasoning', 'step_start')),
+  CONSTRAINT message_parts_seq_uniq UNIQUE (message_id, sequence)
+);
+CREATE INDEX IF NOT EXISTS message_parts_msg_seq_idx ON message_parts (message_id, sequence);
+
+-- v1.13.1-B: read-path view. Read sites SELECT FROM messages_with_parts
+-- instead of messages so tool_calls / tool_results / reasoning_parts come
+-- from the granular message_parts table. The COALESCE means pre-v1.13.0
+-- history (no parts rows) still resolves via the legacy JSON columns; the
+-- dual-write from v1.13.0 keeps both in sync for all rows written since.
+-- Writes continue to target `messages` directly — the view is read-only.
+-- Shapes match the in-memory ToolCall / ToolResult types: tool_calls is a
+-- jsonb array of {id, name, args}, tool_results is a single jsonb object
+-- {tool_call_id, output, truncated, error?}. reasoning_parts is new — only
+-- consumed by the inference history fetch (payload.ts) so v1.13.1-C can
+-- wire reasoning into the model payload. Not surfaced in external APIs yet.
+CREATE OR REPLACE VIEW messages_with_parts AS
+SELECT
+  m.id, m.session_id, m.chat_id, m.role, m.content, m.kind, m.status,
+  m.last_seq, m.tokens_used, m.ctx_used, m.ctx_max,
+  m.started_at, m.finished_at, m.created_at, m.metadata,
+  m.summary, m.tail_start_id, m.compacted_at,
+  COALESCE(
+    (SELECT jsonb_agg(p.payload ORDER BY p.sequence)
+       FROM message_parts p
+      WHERE p.message_id = m.id AND p.kind = 'tool_call'),
+    m.tool_calls
+  ) AS tool_calls,
+  COALESCE(
+    (SELECT p.payload
+       FROM message_parts p
+      WHERE p.message_id = m.id AND p.kind = 'tool_result'
+      ORDER BY p.sequence
+      LIMIT 1),
+    m.tool_results
+  ) AS tool_results,
+  (SELECT jsonb_agg(p.payload ORDER BY p.sequence)
+     FROM message_parts p
+    WHERE p.message_id = m.id AND p.kind = 'reasoning') AS reasoning_parts
+FROM messages m;
+
 ALTER TABLE messages ADD COLUMN IF NOT EXISTS tokens_used INTEGER;
 ALTER TABLE messages ADD COLUMN IF NOT EXISTS ctx_used INTEGER;
 ALTER TABLE messages ADD COLUMN IF NOT EXISTS ctx_max INTEGER;

apps/server/src/services/__tests__/doom-loop.test.ts

@@ -1,5 +1,5 @@
 import { describe, it, expect } from 'vitest';
-import { DOOM_LOOP_THRESHOLD, detectDoomLoop } from '../inference.js';
+import { DOOM_LOOP_THRESHOLD, detectDoomLoop } from '../inference/index.js';
 import type { ToolCall } from '../../types/api.js';
 
 // ---- fixture ----------------------------------------------------------------

apps/server/src/services/__tests__/inference.test.ts

@@ -1,5 +1,5 @@
 import { describe, it, expect } from 'vitest';
-import { buildMessagesPayload } from '../inference.js';
+import { buildMessagesPayload } from '../inference/index.js';
 import type {
   Message,
   MessageRole,

apps/server/src/services/__tests__/parts.test.ts

@@ -0,0 +1,121 @@
+import { describe, it, expect } from 'vitest';
+import { partsFromAssistantMessage, partsFromToolMessage } from '../inference/parts.js';
+import type { ToolCall, ToolResult } from '../../types/api.js';
+
+describe('partsFromAssistantMessage', () => {
+  it('emits one text part for content-only assistant', () => {
+    const parts = partsFromAssistantMessage({ content: 'hello world', tool_calls: null });
+    expect(parts).toHaveLength(1);
+    expect(parts[0]).toEqual({
+      sequence: 0,
+      kind: 'text',
+      payload: { text: 'hello world' },
+    });
+  });
+
+  it('emits one tool_call part for empty-content + single tool_call', () => {
+    const tc: ToolCall = { id: 'call_1', name: 'view_file', args: { path: 'src/a.ts' } };
+    const parts = partsFromAssistantMessage({ content: '', tool_calls: [tc] });
+    expect(parts).toHaveLength(1);
+    expect(parts[0]).toEqual({
+      sequence: 0,
+      kind: 'tool_call',
+      payload: { id: 'call_1', name: 'view_file', args: { path: 'src/a.ts' } },
+    });
+  });
+
+  it('emits text then tool_call parts in order when both present', () => {
+    const tc: ToolCall = { id: 'call_2', name: 'grep', args: { pattern: 'foo' } };
+    const parts = partsFromAssistantMessage({ content: 'let me search', tool_calls: [tc] });
+    expect(parts.map((p) => [p.sequence, p.kind])).toEqual([
+      [0, 'text'],
+      [1, 'tool_call'],
+    ]);
+  });
+
+  it('preserves tool_call order with multiple calls', () => {
+    const calls: ToolCall[] = [
+      { id: 'a', name: 'list_dir', args: { path: '.' } },
+      { id: 'b', name: 'view_file', args: { path: 'x.ts' } },
+      { id: 'c', name: 'grep', args: { pattern: 'y' } },
+    ];
+    const parts = partsFromAssistantMessage({ content: '', tool_calls: calls });
+    expect(parts).toHaveLength(3);
+    expect(parts.map((p) => p.payload)).toEqual([
+      { id: 'a', name: 'list_dir', args: { path: '.' } },
+      { id: 'b', name: 'view_file', args: { path: 'x.ts' } },
+      { id: 'c', name: 'grep', args: { pattern: 'y' } },
+    ]);
+    expect(parts.map((p) => p.sequence)).toEqual([0, 1, 2]);
+  });
+
+  it('returns empty array for empty content + null tool_calls', () => {
+    expect(partsFromAssistantMessage({ content: '', tool_calls: null })).toEqual([]);
+  });
+
+  it('v1.13.1-C: reasoning lands at sequence 0 before text + tool_calls', () => {
+    const tc: ToolCall = { id: 'call_r', name: 'view_file', args: { path: 'x.ts' } };
+    const parts = partsFromAssistantMessage({
+      content: 'inspecting now',
+      tool_calls: [tc],
+      reasoning: 'user asked about x.ts; I should view it',
+    });
+    expect(parts.map((p) => [p.sequence, p.kind])).toEqual([
+      [0, 'reasoning'],
+      [1, 'text'],
+      [2, 'tool_call'],
+    ]);
+    expect(parts[0]!.payload).toEqual({
+      text: 'user asked about x.ts; I should view it',
+    });
+  });
+
+  it('v1.13.1-C: reasoning + empty content + tool_calls preserves seq 0 reasoning', () => {
+    const tc: ToolCall = { id: 'call_r2', name: 'grep', args: { pattern: 'foo' } };
+    const parts = partsFromAssistantMessage({
+      content: '',
+      tool_calls: [tc],
+      reasoning: 'jumping straight to grep',
+    });
+    expect(parts.map((p) => [p.sequence, p.kind])).toEqual([
+      [0, 'reasoning'],
+      [1, 'tool_call'],
+    ]);
+  });
+});
+
+describe('partsFromToolMessage', () => {
+  it('emits a single tool_result part at sequence 0', () => {
+    const tr: ToolResult = {
+      tool_call_id: 'call_1',
+      output: { contents: 'console.log(1)' },
+      truncated: false,
+    };
+    const parts = partsFromToolMessage({ tool_results: tr });
+    expect(parts).toHaveLength(1);
+    expect(parts[0]).toEqual({
+      sequence: 0,
+      kind: 'tool_result',
+      payload: {
+        tool_call_id: 'call_1',
+        output: { contents: 'console.log(1)' },
+        truncated: false,
+      },
+    });
+  });
+
+  it('includes error in payload when present', () => {
+    const tr: ToolResult = {
+      tool_call_id: 'call_2',
+      output: null,
+      truncated: false,
+      error: 'permission denied',
+    };
+    const parts = partsFromToolMessage({ tool_results: tr });
+    expect(parts[0]!.payload).toMatchObject({ error: 'permission denied' });
+  });
+
+  it('returns empty array when tool_results is null', () => {
+    expect(partsFromToolMessage({ tool_results: null })).toEqual([]);
+  });
+});

apps/server/src/services/__tests__/tools.test.ts

@@ -0,0 +1,14 @@
+import { describe, it, expect } from 'vitest';
+import { ALL_TOOLS } from '../tools.js';
+
+describe('ALL_TOOLS registry', () => {
+  // v1.13.3: tools must be alpha-sorted at module load. llama.cpp's prompt
+  // cache hits on byte-identical prefixes; the tool list lives near the
+  // top of the system prompt, so any order drift invalidates every cached
+  // turn. The registry sort is the single source of truth; downstream
+  // helpers (toolJsonSchemas, TOOLS_BY_NAME, buildAiTools) inherit it.
+  it('exports tools in alphabetical order by name', () => {
+    const names = ALL_TOOLS.map((t) => t.name);
+    expect(names).toEqual([...names].sort((a, b) => a.localeCompare(b)));
+  });
+});

apps/server/src/services/auto_name.ts

@@ -1,4 +1,4 @@
-import type { InferenceContext } from './inference.js';
+import type { InferenceContext } from './inference/index.js';
 
 const NAMING_SYSTEM_PROMPT =
   'You name chat sessions. Reply directly with no thinking, reasoning, or explanation. Output ONLY the title, 4 words max, no quotes, no punctuation, no prefix like "Title:".';

apps/server/src/services/compaction.ts

@@ -342,9 +342,11 @@ export async function process(input: ProcessInput): Promise<void> {
   // 2. All currently-active messages in this chat (compacted_at IS NULL).
   // ORDER BY (created_at, id) matches loadContext in inference.ts so the
   // turns() boundary logic sees the same sequence the LLM will.
+  // v1.13.1-B: reads tool_calls/tool_results via the parts-merged view so
+  // the compaction payload matches what the LLM saw on the original turn.
   const messages = await sql<CompactionMessage[]>`
     SELECT id, role, content, kind, summary, status, tool_calls, tool_results, metadata, created_at
-    FROM messages
+    FROM messages_with_parts
     WHERE chat_id = ${chatId} AND compacted_at IS NULL
     ORDER BY created_at ASC, id ASC
   `;

apps/server/src/services/inference/budget.ts

@@ -0,0 +1,20 @@
+import type { Agent } from '../../types/api.js';
+import { READ_ONLY_TOOL_NAMES } from '../tools.js';
+
+// v1.8.2: tool-call budget defaults. Resolved per-turn by resolveToolBudget.
+//   - Agent with explicit max_tool_calls: that value.
+//   - Agent with read-only-only tools:    BUDGET_READ_ONLY (30).
+//   - Agent with any non-read-only tool:  BUDGET_NON_READ_ONLY (10).
+//   - No agent (raw chat):                BUDGET_NO_AGENT (15).
+export const BUDGET_READ_ONLY = 30;
+export const BUDGET_NON_READ_ONLY = 10;
+export const BUDGET_NO_AGENT = 15;
+
+const READ_ONLY_SET: ReadonlySet<string> = new Set(READ_ONLY_TOOL_NAMES);
+
+export function resolveToolBudget(agent: Agent | null): number {
+  if (agent?.max_tool_calls != null) return agent.max_tool_calls;
+  if (!agent) return BUDGET_NO_AGENT;
+  const allReadOnly = agent.tools.every((t) => READ_ONLY_SET.has(t));
+  return allReadOnly ? BUDGET_READ_ONLY : BUDGET_NON_READ_ONLY;
+}

apps/server/src/services/inference/error-handler.ts

@@ -0,0 +1,167 @@
+import type { MessageMetadata, Session } from '../../types/api.js';
+import * as modelContext from '../model-context.js';
+import { maybeFlagForCompaction } from './payload.js';
+import { insertParts, partsFromAssistantMessage } from './parts.js';
+import type { InferenceContext, StreamResult, TurnArgs } from './turn.js';
+
+export async function handleAbortOrError(
+  ctx: InferenceContext,
+  args: TurnArgs,
+  accumulated: string,
+  err: unknown
+): Promise<void> {
+  const { sessionId, chatId, assistantMessageId } = args;
+  const isAbort = err instanceof Error && err.name === 'AbortError';
+  const finalStatus = isAbort ? 'cancelled' : 'failed';
+  const errMsg = err instanceof Error ? err.message : String(err);
+  // v1.8.2: persist a structured error metadata blob on genuine failures so
+  // the bubble can render the reason on reload without re-deriving from the
+  // (one-shot) WS error frame. User-initiated abort skips this — there's no
+  // "reason" to surface for a stop the user already explicitly chose.
+  const errorMetadata: MessageMetadata | null = isAbort
+    ? null
+    : { kind: 'error', error_reason: 'llm_provider_error', error_text: errMsg };
+  if (errorMetadata) {
+    await ctx.sql`
+      UPDATE messages
+      SET status = ${finalStatus},
+          content = ${accumulated},
+          finished_at = clock_timestamp(),
+          metadata = ${ctx.sql.json(errorMetadata as never)}
+      WHERE id = ${assistantMessageId}
+    `;
+  } else {
+    await ctx.sql`
+      UPDATE messages
+      SET status = ${finalStatus},
+          content = ${accumulated},
+          finished_at = clock_timestamp()
+      WHERE id = ${assistantMessageId}
+    `;
+  }
+  const [failSessRow] = await ctx.sql<{ project_id: string; name: string; updated_at: string }[]>`
+    UPDATE sessions SET updated_at = clock_timestamp()
+    WHERE id = ${sessionId}
+    RETURNING project_id, name, updated_at
+  `;
+  ctx.publishUser({ type: 'session_updated', session_id: sessionId, project_id: failSessRow!.project_id, name: failSessRow!.name, updated_at: failSessRow!.updated_at });
+  // v1.8 mobile-tabs: cancellation is a user-initiated stop, treat as idle;
+  // genuine errors flip the dot red. v1.8.2: error path also carries a
+  // machine-readable `reason` so the UI can render specifics inline.
+  if (isAbort) {
+    // v1.12.1: defensive cancellation write. The status=${finalStatus} UPDATE
+    // above already sets 'cancelled' for the AbortError case, but a row can
+    // leak as 'streaming' when the abort fires between the post-tool-phase
+    // INSERT (executeToolPhase) and the next runAssistantTurn's stream setup,
+    // bypassing the try/catch around executeStreamPhase. The status guard
+    // makes this a no-op when the earlier write already landed.
+    await ctx.sql`
+      UPDATE messages
+      SET status = 'cancelled', content = ${accumulated}, finished_at = clock_timestamp()
+      WHERE id = ${args.assistantMessageId} AND status = 'streaming'
+    `;
+    ctx.publishUser({ type: 'chat_status', chat_id: chatId, status: 'idle', at: new Date().toISOString() });
+    ctx.publish(sessionId, {
+      type: 'message_complete',
+      message_id: assistantMessageId,
+      chat_id: chatId,
+    });
+    ctx.log.info({ sessionId, chatId, assistantMessageId }, 'inference cancelled');
+  } else {
+    ctx.publishUser({
+      type: 'chat_status',
+      chat_id: chatId,
+      status: 'error',
+      at: new Date().toISOString(),
+      reason: 'llm_provider_error',
+    });
+    ctx.publish(sessionId, {
+      type: 'error',
+      message_id: assistantMessageId,
+      chat_id: chatId,
+      error: errMsg,
+      reason: 'llm_provider_error',
+    });
+    ctx.log.error({ err, sessionId, assistantMessageId }, 'inference failed');
+  }
+}
+
+export async function finalizeCompletion(
+  ctx: InferenceContext,
+  args: TurnArgs,
+  result: StreamResult,
+  startedAt: string | null,
+  session: Session
+): Promise<void> {
+  const { sessionId, chatId, assistantMessageId } = args;
+  const { content, finishReason, promptTokens, completionTokens } = result;
+
+  // v1.11.3: see executeToolPhase for the rationale.
+  const mctx = await modelContext.getModelContext(session.model);
+  const nCtx = mctx?.n_ctx ?? null;
+
+  const [updated] = await ctx.sql<
+    { tokens_used: number | null; ctx_used: number | null; ctx_max: number | null; finished_at: string | null }[]
+  >`
+    UPDATE messages
+    SET content = ${content},
+        status = 'complete',
+        tokens_used = ${completionTokens},
+        ctx_used = ${promptTokens},
+        ctx_max = ${nCtx},
+        finished_at = clock_timestamp()
+    WHERE id = ${assistantMessageId}
+    RETURNING tokens_used, ctx_used, ctx_max, finished_at
+  `;
+  // v1.13.0: dual-write the text part. finalizeCompletion is the terminal
+  // path for text-only assistant turns (no tool calls); tool_calls are null
+  // here by construction (the tool-bearing path goes through executeToolPhase).
+  // v1.13.1-C: include result.reasoning so reasoning-channel models capture
+  // a kind='reasoning' part alongside the text.
+  // TODO(v1.13.1): wrap the UPDATE above and this insertParts in a single
+  // sql.begin before flipping read authority to message_parts.
+  await insertParts(
+    ctx.sql,
+    partsFromAssistantMessage({
+      content,
+      tool_calls: null,
+      reasoning: result.reasoning,
+    }).map((p) => ({
+      ...p,
+      message_id: assistantMessageId,
+    })),
+  );
+  // v1.11: flag for compaction on the terminal turn too. Catches the common
+  // case of a turn that hit the limit without invoking tools.
+  await maybeFlagForCompaction(ctx, chatId, updated);
+  const [completeSessRow] = await ctx.sql<{ project_id: string; name: string; updated_at: string }[]>`
+    UPDATE sessions SET updated_at = clock_timestamp()
+    WHERE id = ${sessionId}
+    RETURNING project_id, name, updated_at
+  `;
+  ctx.publishUser({ type: 'session_updated', session_id: sessionId, project_id: completeSessRow!.project_id, name: completeSessRow!.name, updated_at: completeSessRow!.updated_at });
+  ctx.publishUser({ type: 'chat_status', chat_id: chatId, status: 'idle', at: new Date().toISOString() });
+  ctx.publish(sessionId, {
+    type: 'message_complete',
+    message_id: assistantMessageId,
+    chat_id: chatId,
+    tokens_used: updated?.tokens_used ?? null,
+    ctx_used: updated?.ctx_used ?? null,
+    ctx_max: updated?.ctx_max ?? null,
+    started_at: startedAt,
+    finished_at: updated?.finished_at ?? null,
+    model: session.model,
+  });
+  ctx.log.info(
+    {
+      sessionId,
+      chatId,
+      assistantMessageId,
+      finishReason,
+      chars: content.length,
+      tokens_used: updated?.tokens_used,
+      ctx_used: updated?.ctx_used,
+    },
+    'inference complete'
+  );
+}

apps/server/src/services/inference/index.ts

@@ -0,0 +1,20 @@
+// v1.12.4: re-export shim. Outside callers (apps/server/src/index.ts and the
+// vitest inference tests) import from './services/inference/index.js'. The
+// directory is now the public surface; turn.ts holds runAssistantTurn /
+// runInference / createInferenceRunner while the other inference/*.ts files
+// stay implementation-private.
+
+export {
+  createInferenceRunner,
+  runAssistantTurn,
+  runInference,
+} from './turn.js';
+export type {
+  FramePublisher,
+  InferenceContext,
+  InferenceFrame,
+  StreamResult,
+  TurnArgs,
+} from './turn.js';
+export { detectDoomLoop, DOOM_LOOP_THRESHOLD } from './sentinels.js';
+export { buildMessagesPayload } from './payload.js';

apps/server/src/services/inference/parts.ts

@@ -0,0 +1,95 @@
+import type { Sql } from '../../db.js';
+import type { ToolCall, ToolResult } from '../../types/api.js';
+
+// v1.13.0: dual-write helper. Every site that writes the legacy
+// messages.tool_calls / messages.tool_results JSON columns calls into here
+// to mirror the same data into message_parts rows. Reads still go to the
+// JSON columns; the swap to parts-as-source-of-truth happens in a later
+// v1.13 dispatch alongside the AI SDK streamText migration.
+
+export type PartKind = 'text' | 'tool_call' | 'tool_result' | 'reasoning' | 'step_start';
+
+export interface PartInsert {
+  message_id: string;
+  sequence: number;
+  kind: PartKind;
+  payload: unknown;
+}
+
+export async function insertParts(sql: Sql, parts: PartInsert[]): Promise<void> {
+  if (parts.length === 0) return;
+  // postgres-js fans out an array of objects to a multi-row INSERT. Each
+  // payload field needs sql.json() so jsonb storage receives a JSON value
+  // rather than a quoted string.
+  await sql`
+    INSERT INTO message_parts ${sql(
+      parts.map((p) => ({
+        message_id: p.message_id,
+        sequence: p.sequence,
+        kind: p.kind,
+        payload: sql.json(p.payload as never),
+      })),
+      'message_id',
+      'sequence',
+      'kind',
+      'payload',
+    )}
+  `;
+}
+
+// Derive parts from the canonical messages row for an assistant message.
+// reasoning (when non-empty) becomes a 'reasoning' part at sequence 0 —
+// it precedes user-visible content logically. content (when non-empty)
+// becomes a 'text' part next; each tool_call becomes a 'tool_call' part
+// with payload { id, name, args } where args is the parsed object (we
+// use the in-memory ToolCall shape, not the OpenAI stringified one).
+export function partsFromAssistantMessage(args: {
+  content: string;
+  tool_calls: ToolCall[] | null;
+  // v1.13.1-C: optional reasoning text streamed alongside the answer.
+  // Most rows have none — only models with separate reasoning channels
+  // (qwen3.6 etc.) populate this.
+  reasoning?: string;
+}): Omit<PartInsert, 'message_id'>[] {
+  const out: Omit<PartInsert, 'message_id'>[] = [];
+  let seq = 0;
+  if (args.reasoning && args.reasoning.length > 0) {
+    out.push({ sequence: seq, kind: 'reasoning', payload: { text: args.reasoning } });
+    seq += 1;
+  }
+  if (args.content && args.content.length > 0) {
+    out.push({ sequence: seq, kind: 'text', payload: { text: args.content } });
+    seq += 1;
+  }
+  for (const tc of args.tool_calls ?? []) {
+    out.push({
+      sequence: seq,
+      kind: 'tool_call',
+      payload: { id: tc.id, name: tc.name, args: tc.args },
+    });
+    seq += 1;
+  }
+  return out;
+}
+
+// Derive a single tool_result part from a tool message's tool_results JSON.
+// The payload includes the same shape that buildMessagesPayload reads from
+// later: tool_call_id, output, optional error/truncated metadata.
+export function partsFromToolMessage(args: {
+  tool_results: ToolResult | null;
+}): Omit<PartInsert, 'message_id'>[] {
+  if (!args.tool_results) return [];
+  const tr = args.tool_results;
+  return [
+    {
+      sequence: 0,
+      kind: 'tool_result',
+      payload: {
+        tool_call_id: tr.tool_call_id,
+        output: tr.output,
+        truncated: tr.truncated,
+        ...(tr.error ? { error: tr.error } : {}),
+      },
+    },
+  ];
+}

apps/server/src/services/inference/payload.ts

@@ -0,0 +1,171 @@
+import type { Sql } from '../../db.js';
+import type {
+  Agent,
+  Message,
+  Project,
+  Session,
+} from '../../types/api.js';
+import * as compaction from '../compaction.js';
+import { buildSystemPrompt } from '../system-prompt.js';
+import { isAnySentinel } from './sentinels.js';
+import type { InferenceContext } from './turn.js';
+
+export interface OpenAiMessage {
+  role: 'system' | 'user' | 'assistant' | 'tool';
+  content: string | null;
+  tool_calls?: Array<{
+    id: string;
+    type: 'function';
+    function: { name: string; arguments: string };
+  }>;
+  tool_call_id?: string;
+  // v1.13.1-C: reasoning text from a prior assistant turn, sourced from
+  // message_parts kind='reasoning' rows joined in via reasoning_parts on
+  // the messages_with_parts view. stream-phase.ts/toModelMessages threads
+  // this into the AI SDK ReasoningPart when forwarding to the model so
+  // reasoning models can resume mid-thought across tool-call boundaries.
+  reasoning?: string;
+}
+
+// v1.12: buildSystemPrompt lives in services/system-prompt.ts. It awaits the
+// container-guidance loader, so this function is async too and every call
+// site in inference.ts awaits the result.
+export async function buildMessagesPayload(
+  session: Session,
+  project: Project,
+  history: Message[],
+  agent: Agent | null = null
+): Promise<OpenAiMessage[]> {
+  const out: OpenAiMessage[] = [];
+  const systemPrompt = await buildSystemPrompt(project, session, agent);
+  out.push({ role: 'system', content: systemPrompt });
+
+  // Find the latest compact marker — only send messages from that point onwards
+  let startIdx = 0;
+  for (let i = history.length - 1; i >= 0; i--) {
+    if (history[i]!.kind === 'compact') {
+      startIdx = i;
+      break;
+    }
+  }
+
+  for (let i = startIdx; i < history.length; i++) {
+    const m = history[i]!;
+    if (m.kind === 'compact') {
+      out.push({ role: 'system', content: m.content });
+      continue;
+    }
+    // v1.8.2 / v1.11.6: cap-hit and doom-loop sentinels are UI-only — never
+    // send them to the LLM. The synthetic instruction note lives only inside
+    // the summary call's messages array and is never persisted, so on a
+    // follow-up turn the model resumes with a clean context.
+    if (isAnySentinel(m)) continue;
+    if (m.role === 'assistant' && m.status === 'streaming') continue;
+    if (m.role === 'assistant' && m.status === 'cancelled') continue;
+    if (m.role === 'tool') {
+      const tr = m.tool_results;
+      if (!tr) continue;
+      const outputText = tr.error
+        ? `error: ${tr.error}`
+        : typeof tr.output === 'string'
+          ? tr.output
+          : JSON.stringify(tr.output);
+      out.push({
+        role: 'tool',
+        content: outputText,
+        tool_call_id: tr.tool_call_id,
+      });
+      continue;
+    }
+    if (m.role === 'assistant') {
+      const msg: OpenAiMessage = {
+        role: 'assistant',
+        content: m.content && m.content.length > 0 ? m.content : null,
+      };
+      if (m.tool_calls && m.tool_calls.length > 0) {
+        msg.tool_calls = m.tool_calls.map((tc) => ({
+          id: tc.id,
+          type: 'function' as const,
+          function: { name: tc.name, arguments: JSON.stringify(tc.args) },
+        }));
+      }
+      // v1.13.1-C: collapse reasoning_parts into a single string. The view
+      // returns them ordered by sequence; multiple reasoning parts on one
+      // message are rare but concat preserves ordering. Skip when absent.
+      if (m.reasoning_parts && m.reasoning_parts.length > 0) {
+        msg.reasoning = m.reasoning_parts.map((p) => p.text ?? '').join('');
+      }
+      out.push(msg);
+      continue;
+    }
+    out.push({ role: 'user', content: m.content });
+  }
+  return out;
+}
+
+export async function loadContext(
+  sql: Sql,
+  sessionId: string,
+  chatId: string
+): Promise<{ session: Session; project: Project; history: Message[] } | null> {
+  const sessionRows = await sql<Session[]>`
+    SELECT id, project_id, name, model, system_prompt, status, created_at, updated_at,
+           agent_id, web_search_enabled
+    FROM sessions WHERE id = ${sessionId}
+  `;
+  if (sessionRows.length === 0) return null;
+  const session = sessionRows[0]!;
+
+  const projectRows = await sql<Project[]>`
+    SELECT id, name, path, added_at, last_session_id, status, gitea_remote,
+           default_system_prompt, default_web_search_enabled
+    FROM projects WHERE id = ${session.project_id}
+  `;
+  if (projectRows.length === 0) return null;
+  const project = projectRows[0]!;
+
+  // v1.11: filter compacted messages out of the inference assembly. The GET
+  // /api/sessions/:id/messages endpoint still returns everything (so the UI
+  // can show history with the summary card inline); only LLM payloads skip
+  // compacted rows. compacted_at IS NULL keeps the active summary + tail.
+  // v1.13.1-B: reads tool_calls/tool_results via the parts-merged view.
+  // v1.13.1-C: also pull reasoning_parts so assistant messages from
+  // reasoning models can be replayed with their reasoning context preserved.
+  const history = await sql<Message[]>`
+    SELECT id, session_id, chat_id, role, content, kind, tool_calls, tool_results, status, last_seq,
+           tokens_used, ctx_used, ctx_max, started_at, finished_at, created_at, metadata,
+           reasoning_parts
+    FROM messages_with_parts
+    WHERE chat_id = ${chatId} AND compacted_at IS NULL
+    ORDER BY created_at ASC, id ASC
+  `;
+
+  return { session, project, history };
+}
+
+// v1.11: shared helper used after both finalizeCompletion and executeToolPhase
+// persist their token counts. Reads tokens off the just-UPDATEd row (which
+// the caller returns from RETURNING), runs compaction.isOverflow, and flips
+// chats.needs_compaction. The next runAssistantTurn invocation acts on it.
+// Silent on missing tokens — llama-swap occasionally omits usage on truncated
+// streams, and we'd rather miss one overflow than crash the inference path.
+export async function maybeFlagForCompaction(
+  ctx: InferenceContext,
+  chatId: string,
+  updated: { tokens_used: number | null; ctx_used: number | null; ctx_max: number | null } | undefined,
+): Promise<void> {
+  if (!updated) return;
+  const promptTokens = updated.ctx_used;
+  const completionTokens = updated.tokens_used;
+  const contextLimit = updated.ctx_max;
+  if (typeof promptTokens !== 'number') return;
+  if (typeof completionTokens !== 'number') return;
+  if (typeof contextLimit !== 'number') return;
+  const overflow = compaction.isOverflow(
+    { prompt_tokens: promptTokens, completion_tokens: completionTokens },
+    contextLimit,
+  );
+  if (!overflow) return;
+  await ctx.sql`UPDATE chats SET needs_compaction = true WHERE id = ${chatId}`;
+  ctx.log.info({ chatId, promptTokens, completionTokens, contextLimit }, 'inference: flagged for compaction');
+}

apps/server/src/services/inference/provider.ts

@@ -0,0 +1,26 @@
+import { createOpenAICompatible } from '@ai-sdk/openai-compatible';
+import type { LanguageModel } from 'ai';
+
+// v1.13.1-A: AI SDK provider against llama-swap. baseURL is threaded from
+// config.LLAMA_SWAP_URL at call time (not module-load) so tests can stub the
+// upstream without touching env vars. No apiKey — llama-swap is unauth in our
+// Tailscale topology and exposing it over the public internet is gated by
+// Authelia at the Caddy layer, not by API keys.
+
+const cache = new Map<string, ReturnType<typeof createOpenAICompatible>>();
+
+function getProvider(baseURL: string): ReturnType<typeof createOpenAICompatible> {
+  let provider = cache.get(baseURL);
+  if (!provider) {
+    provider = createOpenAICompatible({
+      name: 'llama-swap',
+      baseURL: baseURL.endsWith('/v1') ? baseURL : `${baseURL}/v1`,
+    });
+    cache.set(baseURL, provider);
+  }
+  return provider;
+}
+
+export function upstreamModel(baseURL: string, modelId: string): LanguageModel {
+  return getProvider(baseURL).chatModel(modelId);
+}

apps/server/src/services/inference/sentinel-summaries.ts

@@ -0,0 +1,523 @@
+import type {
+  Agent,
+  Message,
+  MessageMetadata,
+  Project,
+  Session,
+} from '../../types/api.js';
+import * as modelContext from '../model-context.js';
+import { buildMessagesPayload } from './payload.js';
+import { DOOM_LOOP_THRESHOLD } from './sentinels.js';
+import { streamCompletion } from './stream-phase.js';
+import { DB_FLUSH_INTERVAL_MS } from './types.js';
+import type {
+  InferenceContext,
+  StreamResult,
+  TurnArgs,
+} from './turn.js';
+
+// Synthetic system note appended to the cap-hit summary call. Verbatim from
+// the v1.8.2 spec — do not paraphrase: the model is more reliable when the
+// instruction is short, declarative, and identical across calls.
+const CAP_HIT_SUMMARY_NOTE = (limit: number) =>
+  `You've reached the tool budget (${limit} calls). Produce the best answer you can with what you have. Do not call more tools.`;
+
+const DOOM_LOOP_NOTE = (name: string) =>
+  `You called ${name} with the same arguments ${DOOM_LOOP_THRESHOLD} times in a row. Stop calling it. Produce the best answer you can with what you have.`;
+
+export async function runCapHitSummary(
+  ctx: InferenceContext,
+  args: TurnArgs,
+  session: Session,
+  project: Project,
+  history: Message[],
+  agent: Agent | null,
+  budget: number,
+): Promise<void> {
+  const { sessionId, chatId, assistantMessageId, signal } = args;
+
+  const messages = await buildMessagesPayload(session, project, history, agent);
+  messages.push({ role: 'system', content: CAP_HIT_SUMMARY_NOTE(budget) });
+
+  const startedRow = await ctx.sql<{ started_at: string }[]>`
+    UPDATE messages
+    SET started_at = clock_timestamp()
+    WHERE id = ${assistantMessageId}
+    RETURNING started_at
+  `;
+  const startedAt = startedRow[0]?.started_at ?? null;
+
+  ctx.publish(sessionId, {
+    type: 'message_started',
+    message_id: assistantMessageId,
+    chat_id: chatId,
+    role: 'assistant',
+  });
+
+  let accumulated = '';
+  let pendingFlushTimer: NodeJS.Timeout | null = null;
+  let flushPromise: Promise<unknown> = Promise.resolve();
+  const flushNow = () => {
+    if (pendingFlushTimer) {
+      clearTimeout(pendingFlushTimer);
+      pendingFlushTimer = null;
+    }
+    const snapshot = accumulated;
+    flushPromise = flushPromise.then(() =>
+      ctx.sql`UPDATE messages SET content = ${snapshot} WHERE id = ${assistantMessageId}`
+    );
+  };
+  const scheduleFlush = () => {
+    if (pendingFlushTimer) return;
+    pendingFlushTimer = setTimeout(() => {
+      pendingFlushTimer = null;
+      flushNow();
+    }, DB_FLUSH_INTERVAL_MS);
+  };
+
+  let summaryOk = false;
+  let summarySoftCancelled = false;
+  let summaryError: string | null = null;
+  let result: StreamResult | null = null;
+  try {
+    result = await streamCompletion(
+      ctx,
+      session.model,
+      messages,
+      { tools: null, temperature: agent?.temperature },
+      (delta) => {
+        accumulated += delta;
+        ctx.publish(sessionId, {
+          type: 'delta',
+          message_id: assistantMessageId,
+          chat_id: chatId,
+          content: delta,
+        });
+        scheduleFlush();
+      },
+      undefined,
+      signal,
+    );
+    summaryOk = true;
+  } catch (err) {
+    if (err instanceof Error && err.name === 'AbortError') {
+      summarySoftCancelled = true;
+    } else {
+      summaryError = err instanceof Error ? err.message : String(err);
+    }
+  } finally {
+    if (pendingFlushTimer) {
+      clearTimeout(pendingFlushTimer);
+      pendingFlushTimer = null;
+    }
+    await flushPromise;
+  }
+
+  // Finalize the summary message based on the three outcomes. The sentinel
+  // is inserted regardless so the user always has the Continue affordance —
+  // even on a partial / failed summary the chat history shows where the
+  // budget was hit.
+  if (summaryOk && result) {
+    // v1.11.3: see executeToolPhase for the rationale.
+    const mctx = await modelContext.getModelContext(session.model);
+    const nCtx = mctx?.n_ctx ?? null;
+    const [updated] = await ctx.sql<
+      { tokens_used: number | null; ctx_used: number | null; ctx_max: number | null; finished_at: string | null }[]
+    >`
+      UPDATE messages
+      SET content = ${result.content},
+          status = 'complete',
+          tokens_used = ${result.completionTokens},
+          ctx_used = ${result.promptTokens},
+          ctx_max = ${nCtx},
+          finished_at = clock_timestamp()
+      WHERE id = ${assistantMessageId}
+      RETURNING tokens_used, ctx_used, ctx_max, finished_at
+    `;
+    ctx.publish(sessionId, {
+      type: 'message_complete',
+      message_id: assistantMessageId,
+      chat_id: chatId,
+      tokens_used: updated?.tokens_used ?? null,
+      ctx_used: updated?.ctx_used ?? null,
+      ctx_max: updated?.ctx_max ?? null,
+      started_at: startedAt,
+      finished_at: updated?.finished_at ?? null,
+      model: session.model,
+    });
+  } else if (summarySoftCancelled) {
+    await ctx.sql`
+      UPDATE messages
+      SET content = ${accumulated},
+          status = 'cancelled',
+          finished_at = clock_timestamp()
+      WHERE id = ${assistantMessageId}
+    `;
+    ctx.publish(sessionId, {
+      type: 'message_complete',
+      message_id: assistantMessageId,
+      chat_id: chatId,
+    });
+  } else {
+    const errMeta: MessageMetadata = {
+      kind: 'error',
+      error_reason: 'summary_after_cap_failed',
+      error_text: summaryError ?? 'summary failed',
+    };
+    await ctx.sql`
+      UPDATE messages
+      SET content = ${accumulated},
+          status = 'failed',
+          finished_at = clock_timestamp(),
+          metadata = ${ctx.sql.json(errMeta as never)}
+      WHERE id = ${assistantMessageId}
+    `;
+    ctx.publish(sessionId, {
+      type: 'error',
+      message_id: assistantMessageId,
+      chat_id: chatId,
+      error: summaryError ?? 'summary failed',
+      reason: 'summary_after_cap_failed',
+    });
+  }
+
+  // Bump session/chat updated_at exactly once for this turn.
+  const [sessRow] = await ctx.sql<{ project_id: string; name: string; updated_at: string }[]>`
+    UPDATE sessions SET updated_at = clock_timestamp()
+    WHERE id = ${sessionId}
+    RETURNING project_id, name, updated_at
+  `;
+  ctx.publishUser({
+    type: 'session_updated',
+    session_id: sessionId,
+    project_id: sessRow!.project_id,
+    name: sessRow!.name,
+    updated_at: sessRow!.updated_at,
+  });
+
+  await insertCapHitSentinel(ctx, sessionId, chatId, agent, budget);
+
+  // Status frame fires last so the dot color reflects the terminal state.
+  // Success → idle, abort → idle (user-driven stop), error → error+reason.
+  if (summaryOk) {
+    ctx.publishUser({ type: 'chat_status', chat_id: chatId, status: 'idle', at: new Date().toISOString() });
+  } else if (summarySoftCancelled) {
+    ctx.publishUser({ type: 'chat_status', chat_id: chatId, status: 'idle', at: new Date().toISOString() });
+  } else {
+    ctx.publishUser({
+      type: 'chat_status',
+      chat_id: chatId,
+      status: 'error',
+      at: new Date().toISOString(),
+      reason: 'summary_after_cap_failed',
+    });
+  }
+
+  ctx.log.info(
+    { sessionId, chatId, assistantMessageId, budget, summaryOk, summaryCancelled: summarySoftCancelled },
+    'inference cap-hit summary finished',
+  );
+}
+
+async function insertCapHitSentinel(
+  ctx: InferenceContext,
+  sessionId: string,
+  chatId: string,
+  agent: Agent | null,
+  budget: number,
+): Promise<void> {
+  // Hard ceiling: count prior cap_hit sentinels in this chat. After two
+  // continues (sentinel count of 2), the next sentinel reports can_continue
+  // false and the UI disables the Continue button.
+  const priorRows = await ctx.sql<{ count: number }[]>`
+    SELECT COUNT(*)::int AS count
+    FROM messages
+    WHERE chat_id = ${chatId}
+      AND role = 'system'
+      AND metadata->>'kind' = 'cap_hit'
+  `;
+  const priorCount = priorRows[0]?.count ?? 0;
+  const canContinue = priorCount < 2;
+  const metadata: MessageMetadata = {
+    kind: 'cap_hit',
+    used: budget,
+    limit: budget,
+    agent_name: agent?.name ?? null,
+    can_continue: canContinue,
+  };
+  const content = `Reached tool budget (${budget}/${budget}). Continue to extend.`;
+
+  const [row] = await ctx.sql<{ id: string }[]>`
+    INSERT INTO messages (session_id, chat_id, role, content, status, created_at, metadata)
+    VALUES (${sessionId}, ${chatId}, 'system', ${content}, 'complete', clock_timestamp(), ${ctx.sql.json(metadata as never)})
+    RETURNING id
+  `;
+
+  // The sentinel content is static, but we still walk the standard frame
+  // sequence (started → delta → complete) so useSessionStream's reducer
+  // appends it via the same path it uses for streaming assistant messages.
+  // The delta carries the full text in one chunk.
+  ctx.publish(sessionId, {
+    type: 'message_started',
+    message_id: row!.id,
+    chat_id: chatId,
+    role: 'system',
+  });
+  ctx.publish(sessionId, {
+    type: 'delta',
+    message_id: row!.id,
+    chat_id: chatId,
+    content,
+  });
+  ctx.publish(sessionId, {
+    type: 'message_complete',
+    message_id: row!.id,
+    chat_id: chatId,
+    metadata,
+  });
+}
+
+// v1.11.6: doom-loop wrap-up. Mirrors runCapHitSummary structurally — same
+// in-flight-slot reuse, same tools-disabled streaming-summary call, same
+// post-finalize sentinel insert + chat_status drop. Differences:
+//   - synthetic note text comes from DOOM_LOOP_NOTE (names the looping tool)
+//   - sentinel metadata is { kind: 'doom_loop', tool_name, args, threshold }
+//     and has no Continue affordance (manual retry would just re-loop)
+//   - chat_status error path uses reason: 'doom_loop_summary_failed'
+// Kept as a clone rather than refactored into a shared helper because the
+// two summary paths still differ in error reason + sentinel shape; a third
+// sentinel would justify factoring out runWrapUpSummary(opts).
+export async function runDoomLoopSummary(
+  ctx: InferenceContext,
+  args: TurnArgs,
+  session: Session,
+  project: Project,
+  history: Message[],
+  agent: Agent | null,
+  loop: { name: string; args: Record<string, unknown> },
+): Promise<void> {
+  const { sessionId, chatId, assistantMessageId, signal } = args;
+
+  const messages = await buildMessagesPayload(session, project, history, agent);
+  messages.push({ role: 'system', content: DOOM_LOOP_NOTE(loop.name) });
+
+  const startedRow = await ctx.sql<{ started_at: string }[]>`
+    UPDATE messages
+    SET started_at = clock_timestamp()
+    WHERE id = ${assistantMessageId}
+    RETURNING started_at
+  `;
+  const startedAt = startedRow[0]?.started_at ?? null;
+
+  ctx.publish(sessionId, {
+    type: 'message_started',
+    message_id: assistantMessageId,
+    chat_id: chatId,
+    role: 'assistant',
+  });
+
+  let accumulated = '';
+  let pendingFlushTimer: NodeJS.Timeout | null = null;
+  let flushPromise: Promise<unknown> = Promise.resolve();
+  const flushNow = () => {
+    if (pendingFlushTimer) {
+      clearTimeout(pendingFlushTimer);
+      pendingFlushTimer = null;
+    }
+    const snapshot = accumulated;
+    flushPromise = flushPromise.then(() =>
+      ctx.sql`UPDATE messages SET content = ${snapshot} WHERE id = ${assistantMessageId}`
+    );
+  };
+  const scheduleFlush = () => {
+    if (pendingFlushTimer) return;
+    pendingFlushTimer = setTimeout(() => {
+      pendingFlushTimer = null;
+      flushNow();
+    }, DB_FLUSH_INTERVAL_MS);
+  };
+
+  let summaryOk = false;
+  let summarySoftCancelled = false;
+  let summaryError: string | null = null;
+  let result: StreamResult | null = null;
+  try {
+    result = await streamCompletion(
+      ctx,
+      session.model,
+      messages,
+      { tools: null, temperature: agent?.temperature },
+      (delta) => {
+        accumulated += delta;
+        ctx.publish(sessionId, {
+          type: 'delta',
+          message_id: assistantMessageId,
+          chat_id: chatId,
+          content: delta,
+        });
+        scheduleFlush();
+      },
+      undefined,
+      signal,
+    );
+    summaryOk = true;
+  } catch (err) {
+    if (err instanceof Error && err.name === 'AbortError') {
+      summarySoftCancelled = true;
+    } else {
+      summaryError = err instanceof Error ? err.message : String(err);
+    }
+  } finally {
+    if (pendingFlushTimer) {
+      clearTimeout(pendingFlushTimer);
+      pendingFlushTimer = null;
+    }
+    await flushPromise;
+  }
+
+  if (summaryOk && result) {
+    const mctx = await modelContext.getModelContext(session.model);
+    const nCtx = mctx?.n_ctx ?? null;
+    const [updated] = await ctx.sql<
+      { tokens_used: number | null; ctx_used: number | null; ctx_max: number | null; finished_at: string | null }[]
+    >`
+      UPDATE messages
+      SET content = ${result.content},
+          status = 'complete',
+          tokens_used = ${result.completionTokens},
+          ctx_used = ${result.promptTokens},
+          ctx_max = ${nCtx},
+          finished_at = clock_timestamp()
+      WHERE id = ${assistantMessageId}
+      RETURNING tokens_used, ctx_used, ctx_max, finished_at
+    `;
+    ctx.publish(sessionId, {
+      type: 'message_complete',
+      message_id: assistantMessageId,
+      chat_id: chatId,
+      tokens_used: updated?.tokens_used ?? null,
+      ctx_used: updated?.ctx_used ?? null,
+      ctx_max: updated?.ctx_max ?? null,
+      started_at: startedAt,
+      finished_at: updated?.finished_at ?? null,
+      model: session.model,
+    });
+  } else if (summarySoftCancelled) {
+    await ctx.sql`
+      UPDATE messages
+      SET content = ${accumulated},
+          status = 'cancelled',
+          finished_at = clock_timestamp()
+      WHERE id = ${assistantMessageId}
+    `;
+    ctx.publish(sessionId, {
+      type: 'message_complete',
+      message_id: assistantMessageId,
+      chat_id: chatId,
+    });
+  } else {
+    // Doom-loop summary failure reuses the existing summary_after_cap_failed
+    // error reason — the ErrorReason union is shared between sentinel paths
+    // and the UI surfaces a generic "summary failed" line for both. We don't
+    // add a new reason code because the user-visible failure mode is the
+    // same (model gave up mid-summary). Sentinel below still fires.
+    const errMeta: MessageMetadata = {
+      kind: 'error',
+      error_reason: 'summary_after_cap_failed',
+      error_text: summaryError ?? 'doom-loop summary failed',
+    };
+    await ctx.sql`
+      UPDATE messages
+      SET content = ${accumulated},
+          status = 'failed',
+          finished_at = clock_timestamp(),
+          metadata = ${ctx.sql.json(errMeta as never)}
+      WHERE id = ${assistantMessageId}
+    `;
+    ctx.publish(sessionId, {
+      type: 'error',
+      message_id: assistantMessageId,
+      chat_id: chatId,
+      error: summaryError ?? 'doom-loop summary failed',
+      reason: 'summary_after_cap_failed',
+    });
+  }
+
+  const [sessRow] = await ctx.sql<{ project_id: string; name: string; updated_at: string }[]>`
+    UPDATE sessions SET updated_at = clock_timestamp()
+    WHERE id = ${sessionId}
+    RETURNING project_id, name, updated_at
+  `;
+  ctx.publishUser({
+    type: 'session_updated',
+    session_id: sessionId,
+    project_id: sessRow!.project_id,
+    name: sessRow!.name,
+    updated_at: sessRow!.updated_at,
+  });
+
+  await insertDoomLoopSentinel(ctx, sessionId, chatId, loop);
+
+  if (summaryOk || summarySoftCancelled) {
+    ctx.publishUser({ type: 'chat_status', chat_id: chatId, status: 'idle', at: new Date().toISOString() });
+  } else {
+    ctx.publishUser({
+      type: 'chat_status',
+      chat_id: chatId,
+      status: 'error',
+      at: new Date().toISOString(),
+      reason: 'summary_after_cap_failed',
+    });
+  }
+
+  ctx.log.info(
+    { sessionId, chatId, assistantMessageId, loopedTool: loop.name, summaryOk, summaryCancelled: summarySoftCancelled },
+    'inference doom-loop summary finished',
+  );
+}
+
+async function insertDoomLoopSentinel(
+  ctx: InferenceContext,
+  sessionId: string,
+  chatId: string,
+  loop: { name: string; args: Record<string, unknown> },
+): Promise<void> {
+  // No hard-ceiling / can-continue logic here — doom-loop is a different
+  // failure mode from cap-hit. Continuing would re-trigger the loop with
+  // the same tools available; the user needs to restate their question
+  // or switch agents instead.
+  const metadata: MessageMetadata = {
+    kind: 'doom_loop',
+    tool_name: loop.name,
+    args: loop.args,
+    threshold: DOOM_LOOP_THRESHOLD,
+  };
+  const content = `Detected ${DOOM_LOOP_THRESHOLD} identical calls to ${loop.name}. Stopping the tool-call loop. Produce the best answer you can with what you have.`;
+
+  const [row] = await ctx.sql<{ id: string }[]>`
+    INSERT INTO messages (session_id, chat_id, role, content, status, created_at, metadata)
+    VALUES (${sessionId}, ${chatId}, 'system', ${content}, 'complete', clock_timestamp(), ${ctx.sql.json(metadata as never)})
+    RETURNING id
+  `;
+
+  // Standard frame sequence — same as cap-hit sentinel — so
+  // useSessionStream's reducer appends the row via the existing path.
+  ctx.publish(sessionId, {
+    type: 'message_started',
+    message_id: row!.id,
+    chat_id: chatId,
+    role: 'system',
+  });
+  ctx.publish(sessionId, {
+    type: 'delta',
+    message_id: row!.id,
+    chat_id: chatId,
+    content,
+  });
+  ctx.publish(sessionId, {
+    type: 'message_complete',
+    message_id: row!.id,
+    chat_id: chatId,
+    metadata,
+  });
+}

apps/server/src/services/inference/sentinels.ts

@@ -0,0 +1,53 @@
+import type { Message, ToolCall } from '../../types/api.js';
+
+// v1.11.6: doom-loop guard. When the model calls the same tool with the
+// same arguments DOOM_LOOP_THRESHOLD times in a row within one user-message
+// turn, abort the recursion and run the same wrap-up summary path as the
+// cap-hit case. Ported from opencode (DOOM_LOOP_THRESHOLD in
+// session/processor.ts). Threshold of 3 is the smallest value that doesn't
+// false-positive on a model that retries once after a transient error.
+export const DOOM_LOOP_THRESHOLD = 3;
+
+// Returns the name + args of the looping tool when the LAST
+// DOOM_LOOP_THRESHOLD entries in `recentToolCalls` are identical (same name
+// AND deep-equal args via JSON.stringify). Returns null otherwise.
+// Pure; exported for unit-test access.
+export function detectDoomLoop(
+  recentToolCalls: ToolCall[],
+): { name: string; args: Record<string, unknown> } | null {
+  if (recentToolCalls.length < DOOM_LOOP_THRESHOLD) return null;
+  const last = recentToolCalls.slice(-DOOM_LOOP_THRESHOLD);
+  const ref = last[0]!;
+  const refArgs = JSON.stringify(ref.args);
+  for (let i = 1; i < last.length; i++) {
+    const tc = last[i]!;
+    if (tc.name !== ref.name) return null;
+    if (JSON.stringify(tc.args) !== refArgs) return null;
+  }
+  return { name: ref.name, args: ref.args };
+}
+
+export function isCapHitSentinel(m: Message): boolean {
+  return (
+    m.role === 'system' &&
+    m.metadata !== null &&
+    typeof m.metadata === 'object' &&
+    (m.metadata as { kind?: unknown }).kind === 'cap_hit'
+  );
+}
+
+// v1.11.6: parallel predicate. Same UI-only semantics as cap-hit sentinels —
+// never sent to the LLM (filtered by buildMessagesPayload through the
+// isAnySentinel check below).
+export function isDoomLoopSentinel(m: Message): boolean {
+  return (
+    m.role === 'system' &&
+    m.metadata !== null &&
+    typeof m.metadata === 'object' &&
+    (m.metadata as { kind?: unknown }).kind === 'doom_loop'
+  );
+}
+
+export function isAnySentinel(m: Message): boolean {
+  return isCapHitSentinel(m) || isDoomLoopSentinel(m);
+}

apps/server/src/services/inference/stream-phase.ts

@@ -0,0 +1,482 @@
+import type {
+  Agent,
+  Session,
+  ToolCall,
+} from '../../types/api.js';
+import * as modelContext from '../model-context.js';
+import { toolJsonSchemas, type ToolJsonSchema } from '../tools.js';
+import type { OpenAiMessage } from './payload.js';
+import {
+  XML_TOOL_CLOSE,
+  XML_TOOL_OPEN,
+  parseXmlToolCall,
+  partialXmlOpenerStart,
+} from './xml-parser.js';
+import { DB_FLUSH_INTERVAL_MS, type StreamPhaseState } from './types.js';
+import type {
+  InferenceContext,
+  StreamResult,
+  TurnArgs,
+} from './turn.js';
+import { upstreamModel } from './provider.js';
+import {
+  jsonSchema,
+  streamText,
+  tool,
+  type JSONValue,
+  type ModelMessage,
+  type ToolCallRepairFunction,
+} from 'ai';
+
+interface StreamOptions {
+  // null = omit tools entirely (compact phase); [] = caller stripped all tools
+  // (rare; we still omit from the request body to avoid OpenAI 400).
+  tools: ToolJsonSchema[] | null;
+  temperature?: number;
+}
+
+// v1.13.1-A: convert BooCode's OpenAI-shaped history into AI SDK
+// ModelMessage[]. Tool result messages need a `toolName` field that the
+// OpenAI shape doesn't carry; we look it up by scanning earlier assistant
+// `tool_calls` entries for a matching id.
+function toModelMessages(messages: OpenAiMessage[]): ModelMessage[] {
+  const toolNameById = new Map<string, string>();
+  for (const m of messages) {
+    if (m.role === 'assistant' && m.tool_calls) {
+      for (const tc of m.tool_calls) {
+        toolNameById.set(tc.id, tc.function.name);
+      }
+    }
+  }
+  const out: ModelMessage[] = [];
+  for (const m of messages) {
+    if (m.role === 'system' || m.role === 'user') {
+      out.push({ role: m.role, content: m.content ?? '' });
+      continue;
+    }
+    if (m.role === 'assistant') {
+      const hasTools = m.tool_calls && m.tool_calls.length > 0;
+      const hasReasoning = typeof m.reasoning === 'string' && m.reasoning.length > 0;
+      if (!hasTools && !hasReasoning) {
+        // Bare text assistant (string content). null content + no tool_calls
+        // is degenerate but harmless to forward.
+        out.push({ role: 'assistant', content: m.content ?? '' });
+        continue;
+      }
+      // v1.13.1-C: AI SDK ReasoningPart precedes text + tool-calls in the
+      // assistant content array. Reasoning models (qwen3.6) consume their
+      // prior reasoning context to resume mid-thought across tool boundaries.
+      const parts: Array<
+        | { type: 'reasoning'; text: string }
+        | { type: 'text'; text: string }
+        | { type: 'tool-call'; toolCallId: string; toolName: string; input: unknown }
+      > = [];
+      if (hasReasoning) {
+        parts.push({ type: 'reasoning', text: m.reasoning! });
+      }
+      if (m.content && m.content.length > 0) {
+        parts.push({ type: 'text', text: m.content });
+      }
+      for (const tc of m.tool_calls ?? []) {
+        let input: unknown = {};
+        try {
+          input = tc.function.arguments.length > 0 ? JSON.parse(tc.function.arguments) : {};
+        } catch {
+          // Malformed args from a prior turn: pass through as a raw blob so
+          // the model sees the same shape it emitted. Wraps the string under
+          // _raw to match the buildMessagesPayload upstream convention.
+          input = { _raw: tc.function.arguments };
+        }
+        parts.push({ type: 'tool-call', toolCallId: tc.id, toolName: tc.function.name, input });
+      }
+      out.push({ role: 'assistant', content: parts });
+      continue;
+    }
+    if (m.role === 'tool') {
+      const toolCallId = m.tool_call_id ?? '';
+      const toolName = toolNameById.get(toolCallId) ?? 'unknown';
+      const raw = m.content ?? '';
+      let output: { type: 'text'; value: string } | { type: 'json'; value: JSONValue };
+      try {
+        // JSON.parse returns `any`; cast to JSONValue since the upstream
+        // tool_results column is already JSON-serializable by construction.
+        output = { type: 'json', value: JSON.parse(raw) as JSONValue };
+      } catch {
+        output = { type: 'text', value: raw };
+      }
+      out.push({
+        role: 'tool',
+        content: [{ type: 'tool-result', toolCallId, toolName, output }],
+      });
+      continue;
+    }
+  }
+  return out;
+}
+
+// Build the AI SDK tools record from BooCode's JSON-schema tool definitions.
+// No `execute` field: BooCode runs tools itself in tool-phase.ts; streamText
+// surfaces the tool-call parts via fullStream and we capture them for the
+// outer loop to dispatch.
+function buildAiTools(schemas: ToolJsonSchema[]): Record<string, ReturnType<typeof tool>> {
+  const out: Record<string, ReturnType<typeof tool>> = {};
+  for (const s of schemas) {
+    out[s.function.name] = tool({
+      description: s.function.description,
+      inputSchema: jsonSchema(s.function.parameters),
+    });
+  }
+  return out;
+}
+
+// v1.10.5 Qwen-coder XML fallback. Some local models (notably qwen3-coder via
+// llama-swap) emit tool calls as inline XML inside delta.content rather than
+// the structured tool_calls field. We extract them out of the streamed text
+// before flushing it to the client, mirroring the pre-AI-SDK behavior.
+//
+// XML shape:
+//   <tool_call>
+//   <function=NAME>
+//   <parameter=KEY>VALUE</parameter>
+//   ...
+//   </function>
+//   </tool_call>
+// Multiple <tool_call> blocks may appear back-to-back; they never nest.
+export async function streamCompletion(
+  ctx: InferenceContext,
+  model: string,
+  messages: OpenAiMessage[],
+  opts: StreamOptions,
+  onDelta: (content: string) => void,
+  onUsage: ((prompt: number | null, completion: number | null) => void) | undefined,
+  signal?: AbortSignal
+): Promise<StreamResult> {
+  const aiMessages = toModelMessages(messages);
+  const hasTools = opts.tools !== null && opts.tools.length > 0;
+  const aiTools = hasTools ? buildAiTools(opts.tools!) : undefined;
+
+  const startedAt = Date.now();
+  // v1.13.1-C: accumulate reasoning text across reasoning-delta parts.
+  // qwen3.6 emits these on a separate channel from text content; we capture
+  // them per stream so finalizeCompletion can dual-write a 'reasoning' part.
+  // Replaces the v1.13.1-A counter-only diagnostic.
+  let reasoningAccumulated = '';
+
+  // v1.13.3: experimental_repairToolCall keeps the stream alive when the
+  // model emits a malformed tool call (bad JSON args, unknown name, etc.).
+  // Without a repair function streamText throws and the WHOLE stream dies;
+  // with one, the SDK invokes us and we route the bad call through normally.
+  // Strategy: pass through unmodified. executeToolPhase's existing error
+  // path (unknown tool name → "unknown tool: X" result; zod-reject → tool
+  // 'X' rejected — fieldname: required) already gives the model a clean
+  // recovery surface on the next turn. Logging gives us visibility into
+  // how often qwen3.6 actually emits broken calls.
+  const repairToolCall: ToolCallRepairFunction<NonNullable<typeof aiTools>> = async ({
+    toolCall,
+    error,
+  }) => {
+    ctx.log.warn(
+      {
+        toolCallId: toolCall.toolCallId,
+        toolName: toolCall.toolName,
+        error: error.message,
+      },
+      'malformed tool call surfaced via repairToolCall',
+    );
+    return toolCall;
+  };
+
+  const result = streamText({
+    model: upstreamModel(ctx.config.LLAMA_SWAP_URL, model),
+    messages: aiMessages,
+    ...(aiTools
+      ? { tools: aiTools, toolChoice: 'auto' as const, experimental_repairToolCall: repairToolCall }
+      : {}),
+    ...(typeof opts.temperature === 'number' ? { temperature: opts.temperature } : {}),
+    abortSignal: signal,
+  });
+
+  let content = '';
+  let pendingBuffer = '';
+  let finishReason: string | null = null;
+  // v1.13.1-A: AI SDK emits one `tool-call` part per fully-aggregated call,
+  // so we no longer need the OpenAI-index reassembly map the manual SSE
+  // parser used. XML tool calls extracted from text content go into the
+  // same flat list and keep the v1.10.5 synthetic id convention.
+  const toolCalls: ToolCall[] = [];
+
+  for await (const part of result.fullStream) {
+    switch (part.type) {
+      case 'text-delta': {
+        pendingBuffer += part.text;
+        // Extract any complete <tool_call>...</tool_call> blocks before
+        // flushing visible text.
+        while (true) {
+          const startIdx = pendingBuffer.indexOf(XML_TOOL_OPEN);
+          if (startIdx === -1) break;
+          const closeIdx = pendingBuffer.indexOf(XML_TOOL_CLOSE, startIdx);
+          if (closeIdx === -1) break;
+          const blockEnd = closeIdx + XML_TOOL_CLOSE.length;
+          const block = pendingBuffer.slice(startIdx, blockEnd);
+          if (startIdx > 0) {
+            const before = pendingBuffer.slice(0, startIdx);
+            content += before;
+            onDelta(before);
+          }
+          const parsedCall = parseXmlToolCall(block);
+          if (parsedCall) {
+            const synthIdx = toolCalls.length;
+            toolCalls.push({
+              id: `xml_call_${synthIdx}`,
+              name: parsedCall.name,
+              args: parsedCall.args,
+            });
+          }
+          // Parse failures still drop the block — leaking <tool_call> XML to
+          // the chat would look worse than silently swallowing the bad block.
+          pendingBuffer = pendingBuffer.slice(blockEnd);
+        }
+        // Hold back any (partial or full) unclosed opener; flush the rest.
+        const partialIdx = partialXmlOpenerStart(pendingBuffer);
+        if (partialIdx >= 0) {
+          if (partialIdx > 0) {
+            const flush = pendingBuffer.slice(0, partialIdx);
+            content += flush;
+            onDelta(flush);
+          }
+          pendingBuffer = pendingBuffer.slice(partialIdx);
+        } else if (pendingBuffer.length > 0) {
+          content += pendingBuffer;
+          onDelta(pendingBuffer);
+          pendingBuffer = '';
+        }
+        break;
+      }
+      case 'tool-call': {
+        // AI SDK has already parsed the input into an object. Match the
+        // ToolCall shape BooCode passes around in toolCallsBuffer downstream.
+        toolCalls.push({
+          id: part.toolCallId,
+          name: part.toolName,
+          args: (part.input ?? {}) as Record<string, unknown>,
+        });
+        break;
+      }
+      case 'reasoning-delta': {
+        // v1.13.1-C: accumulate; finalizeCompletion / executeToolPhase
+        // dual-write the resulting text as a kind='reasoning' part.
+        if (typeof part.text === 'string') {
+          reasoningAccumulated += part.text;
+        }
+        break;
+      }
+      case 'finish': {
+        if (typeof part.finishReason === 'string') {
+          finishReason = part.finishReason;
+        }
+        break;
+      }
+      case 'error': {
+        const err = part.error;
+        throw err instanceof Error ? err : new Error(String(err));
+      }
+      // Intentional no-op: start, start-step, text-start, text-end,
+      // reasoning-start, reasoning-end, source, file, tool-input-start,
+      // tool-input-delta, tool-input-end, tool-result, tool-error,
+      // finish-step, raw. We only care about the aggregated tool-call and
+      // text-delta paths above; the rest are AI SDK lifecycle/streaming
+      // breadcrumbs that don't change BooCode's persistence or WS contract.
+      default:
+        break;
+    }
+  }
+
+  // v1.13.1-A: drain any buffered partial XML opener as plain text. The
+  // pre-AI-SDK path did this on stream end too — better to leak `<tool_c`
+  // than vanish the text.
+  if (pendingBuffer.length > 0) {
+    content += pendingBuffer;
+    onDelta(pendingBuffer);
+    pendingBuffer = '';
+  }
+
+  // AI SDK v6 fullStream returns normally on abort; check signal explicitly.
+  // Without this throw the row would land as status='complete' with partial
+  // content instead of going through handleAbortOrError → status='cancelled'.
+  // Smoke D caught this in v1.13.1-A — don't refactor it away.
+  if (signal?.aborted) {
+    const abortErr = new Error('aborted');
+    abortErr.name = 'AbortError';
+    throw abortErr;
+  }
+
+  // Usage lands as a promise on the result; awaiting after fullStream is
+  // drained is safe. AI SDK v6 names: `inputTokens` / `outputTokens`.
+  let promptTokens: number | null = null;
+  let completionTokens: number | null = null;
+  try {
+    const usage = await result.usage;
+    if (typeof usage.inputTokens === 'number') promptTokens = usage.inputTokens;
+    if (typeof usage.outputTokens === 'number') completionTokens = usage.outputTokens;
+  } catch {
+    // Some providers omit usage on partial streams; leave both null.
+  }
+
+  if (onUsage && (promptTokens !== null || completionTokens !== null)) {
+    onUsage(promptTokens, completionTokens);
+  }
+
+  if (reasoningAccumulated.length > 0) {
+    ctx.log.debug(
+      { reasoningChars: reasoningAccumulated.length, model, elapsed_ms: Date.now() - startedAt },
+      'streamCompletion: captured reasoning',
+    );
+  }
+
+  return {
+    finishReason,
+    content,
+    toolCalls,
+    promptTokens,
+    completionTokens,
+    reasoning: reasoningAccumulated,
+  };
+}
+
+export async function executeStreamPhase(
+  ctx: InferenceContext,
+  args: TurnArgs,
+  session: Session,
+  messages: OpenAiMessage[],
+  state: StreamPhaseState,
+  agent: Agent | null,
+  // v1.11.8: when false, web_search and web_fetch are stripped from the
+  // tool list sent to the LLM, so the model can't even attempt them.
+  webToolsEnabled: boolean,
+): Promise<StreamResult> {
+  const { sessionId, chatId, assistantMessageId, signal } = args;
+
+  const startedRow = await ctx.sql<{ started_at: string }[]>`
+    UPDATE messages
+    SET started_at = clock_timestamp()
+    WHERE id = ${assistantMessageId}
+    RETURNING started_at
+  `;
+  state.startedAt = startedRow[0]?.started_at ?? null;
+
+  ctx.publish(sessionId, {
+    type: 'message_started',
+    message_id: assistantMessageId,
+    chat_id: chatId,
+    role: 'assistant',
+  });
+
+  let pendingFlushTimer: NodeJS.Timeout | null = null;
+  let flushPromise: Promise<unknown> = Promise.resolve();
+
+  const flushNow = () => {
+    if (pendingFlushTimer) {
+      clearTimeout(pendingFlushTimer);
+      pendingFlushTimer = null;
+    }
+    const snapshot = state.accumulated;
+    flushPromise = flushPromise.then(() =>
+      ctx.sql`UPDATE messages SET content = ${snapshot} WHERE id = ${assistantMessageId}`
+    );
+  };
+
+  const scheduleFlush = () => {
+    if (pendingFlushTimer) return;
+    pendingFlushTimer = setTimeout(() => {
+      pendingFlushTimer = null;
+      flushNow();
+    }, DB_FLUSH_INTERVAL_MS);
+  };
+
+  // Tool whitelist: if an agent is set, filter the global tool list to only the
+  // tool names it allows. Unknown names in agent.tools are dropped silently
+  // (handled here by intersection). When no agent: send all tools.
+  // v1.11.8: a second filter strips web_search + web_fetch unless the chat
+  // has them explicitly enabled. Counts as an opt-in security boundary: the
+  // model can't summon a tool that wasn't offered to it.
+  const WEB_TOOL_NAMES: ReadonlySet<string> = new Set(['web_search', 'web_fetch']);
+  const effectiveTools: ToolJsonSchema[] = (agent
+    ? toolJsonSchemas().filter((t) => agent.tools.includes(t.function.name))
+    : toolJsonSchemas()
+  ).filter((t) => webToolsEnabled || !WEB_TOOL_NAMES.has(t.function.name));
+  const effectiveTemperature = agent?.temperature;
+
+  // v1.12.2: ctx_max lookup is cached after the first hit per model, so this
+  // is a Map probe in steady state. We capture nCtx once at the top of the
+  // stream so the throttled usage publish doesn't refetch each tick.
+  const mctxForStream = await modelContext.getModelContext(session.model);
+  const nCtxForStream = mctxForStream?.n_ctx ?? null;
+
+  // v1.12.2 → v1.13.1-A: live usage publishes were throttled to ~500ms when
+  // the manual SSE parser saw `parsed.usage` per chunk. AI SDK v6 surfaces
+  // usage only at stream end (result.usage promise), so the throttle is
+  // effectively a single trailing publish. ChatThroughput will tick once at
+  // stream completion rather than mid-stream — known regression vs v1.12.2,
+  // recovered if a future dispatch interpolates from delta cadence.
+  const USAGE_THROTTLE_MS = 500;
+  let lastUsageAt = 0;
+  let pendingUsage: { p: number | null; c: number | null } | null = null;
+  let usageTimer: NodeJS.Timeout | null = null;
+  const flushUsage = () => {
+    if (!pendingUsage) return;
+    const { p, c } = pendingUsage;
+    pendingUsage = null;
+    lastUsageAt = Date.now();
+    ctx.publish(sessionId, {
+      type: 'usage',
+      message_id: assistantMessageId,
+      chat_id: chatId,
+      completion_tokens: c,
+      ctx_used: p,
+      ctx_max: nCtxForStream,
+    });
+  };
+
+  try {
+    return await streamCompletion(
+      ctx,
+      session.model,
+      messages,
+      { tools: effectiveTools, temperature: effectiveTemperature },
+      (delta) => {
+        state.accumulated += delta;
+        ctx.publish(sessionId, {
+          type: 'delta',
+          message_id: assistantMessageId,
+          chat_id: chatId,
+          content: delta,
+        });
+        ctx.log.debug({ sessionId, delta }, 'inference delta');
+        scheduleFlush();
+      },
+      (prompt, completion) => {
+        pendingUsage = { p: prompt, c: completion };
+        const elapsed = Date.now() - lastUsageAt;
+        if (elapsed >= USAGE_THROTTLE_MS) {
+          flushUsage();
+        } else if (!usageTimer) {
+          usageTimer = setTimeout(() => {
+            usageTimer = null;
+            flushUsage();
+          }, USAGE_THROTTLE_MS - elapsed);
+        }
+      },
+      signal
+    );
+  } finally {
+    if (pendingFlushTimer) {
+      clearTimeout(pendingFlushTimer);
+      pendingFlushTimer = null;
+    }
+    if (usageTimer) {
+      clearTimeout(usageTimer);
+      usageTimer = null;
+    }
+    await flushPromise;
+  }
+}

apps/server/src/services/inference/tool-phase.ts

@@ -0,0 +1,256 @@
+import type { Session, ToolCall } from '../../types/api.js';
+import * as modelContext from '../model-context.js';
+import { PathScopeError } from '../path_guard.js';
+import { TOOLS_BY_NAME } from '../tools.js';
+import { maybeFlagForCompaction } from './payload.js';
+import { insertParts, partsFromAssistantMessage, partsFromToolMessage } from './parts.js';
+import type {
+  InferenceContext,
+  StreamResult,
+  TurnArgs,
+} from './turn.js';
+// v1.12.4: ESM value-import cycle. executeToolPhase recurses into
+// runAssistantTurn which lives in inference.ts. The cycle is safe because
+// the reference is read at call time (inside an async function body), not
+// at module top-level. Node + tsc resolve this cleanly.
+import { runAssistantTurn } from './turn.js';
+
+async function executeToolCall(
+  projectRoot: string,
+  toolCall: ToolCall
+): Promise<{ output: unknown; truncated: boolean; error?: string }> {
+  const tool = TOOLS_BY_NAME[toolCall.name];
+  if (!tool) {
+    return { output: null, truncated: false, error: `unknown tool: ${toolCall.name}` };
+  }
+  const parsed = tool.inputSchema.safeParse(toolCall.args);
+  if (!parsed.success) {
+    // v1.12 Track B.2: enrich the zod-reject path so the model sees a
+    // one-line, tool-named hint ("tool 'search_symbols' rejected — query:
+    // Required") instead of a JSON blob of flatten output. Higher recovery
+    // rate on the next turn; doom-loop guard still bounds infinite retries.
+    // The cast is because tool.inputSchema is ZodType<unknown>, so zod can't
+    // statically narrow flatten()'s fieldErrors key set — but the runtime
+    // shape is the standard { formErrors: string[]; fieldErrors: Record<...> }.
+    const flatten = parsed.error.flatten() as {
+      formErrors: string[];
+      fieldErrors: Record<string, string[] | undefined>;
+    };
+    const fieldErrors = Object.entries(flatten.fieldErrors)
+      .map(([field, errs]) => `${field}: ${errs?.[0] ?? 'invalid'}`)
+      .join('; ');
+    const formError = flatten.formErrors[0];
+    const hint = fieldErrors || formError || 'unknown validation error';
+    return {
+      output: null,
+      truncated: false,
+      error: `tool '${toolCall.name}' rejected — ${hint}`,
+    };
+  }
+  try {
+    const output = await tool.execute(parsed.data, projectRoot);
+    const truncated =
+      typeof output === 'object' && output !== null && 'truncated' in output
+        ? Boolean((output as { truncated: unknown }).truncated)
+        : false;
+    return { output, truncated };
+  } catch (err) {
+    if (err instanceof PathScopeError) {
+      return { output: null, truncated: false, error: err.message };
+    }
+    return {
+      output: null,
+      truncated: false,
+      error: err instanceof Error ? err.message : String(err),
+    };
+  }
+}
+
+export async function executeToolPhase(
+  ctx: InferenceContext,
+  args: TurnArgs,
+  result: StreamResult,
+  startedAt: string | null,
+  session: Session,
+  projectRoot: string
+): Promise<void> {
+  const { sessionId, chatId, assistantMessageId, toolsUsed, signal } = args;
+  const { content, toolCalls, promptTokens, completionTokens } = result;
+
+  // v1.11.3: ctx_max comes from llama-swap /upstream/<model>/props, not the
+  // streaming completion (which doesn't emit n_ctx). getModelContext caches
+  // the positive lookup for the process lifetime, so this is a single Map
+  // hit after the first invocation per model.
+  const mctx = await modelContext.getModelContext(session.model);
+  const nCtx = mctx?.n_ctx ?? null;
+
+  const [updated] = await ctx.sql<
+    { tokens_used: number | null; ctx_used: number | null; ctx_max: number | null; finished_at: string | null }[]
+  >`
+    UPDATE messages
+    SET content = ${content},
+        status = 'complete',
+        tool_calls = ${ctx.sql.json(toolCalls as never)},
+        tokens_used = ${completionTokens},
+        ctx_used = ${promptTokens},
+        ctx_max = ${nCtx},
+        finished_at = clock_timestamp()
+    WHERE id = ${assistantMessageId}
+    RETURNING tokens_used, ctx_used, ctx_max, finished_at
+  `;
+  // v1.13.0: dual-write to message_parts. v1.13.1-B made parts authoritative
+  // for reads via the messages_with_parts view; the JSON column write above
+  // remains for v1.13.1 fallback compatibility (dropped in v1.13.2).
+  // v1.13.1-C: include result.reasoning so models with separate reasoning
+  // channels (qwen3.6) get a kind='reasoning' part at sequence 0.
+  // TODO(v1.13.1): wrap the UPDATE above and this insertParts in a single
+  // sql.begin before flipping read authority to message_parts. Without the
+  // transaction, a crash between the two leaves an orphan message that
+  // becomes invisible in the parts-authoritative read path.
+  await insertParts(
+    ctx.sql,
+    partsFromAssistantMessage({
+      content,
+      tool_calls: toolCalls,
+      reasoning: result.reasoning,
+    }).map((p) => ({
+      ...p,
+      message_id: assistantMessageId,
+    })),
+  );
+  // v1.11: flag for compaction if this turn pushed us over the usable budget.
+  // We never compact mid-loop (the recursive runAssistantTurn keeps tools
+  // flowing); the flag fires on the NEXT turn's pre-fetch hook above.
+  await maybeFlagForCompaction(ctx, chatId, updated);
+  const [toolSessRow] = await ctx.sql<{ project_id: string; name: string; updated_at: string }[]>`
+    UPDATE sessions SET updated_at = clock_timestamp()
+    WHERE id = ${sessionId}
+    RETURNING project_id, name, updated_at
+  `;
+  ctx.publishUser({ type: 'session_updated', session_id: sessionId, project_id: toolSessRow!.project_id, name: toolSessRow!.name, updated_at: toolSessRow!.updated_at });
+  for (const tc of toolCalls) {
+    ctx.publish(sessionId, {
+      type: 'tool_call',
+      message_id: assistantMessageId,
+      chat_id: chatId,
+      tool_call: tc,
+    });
+  }
+  ctx.publish(sessionId, {
+    type: 'message_complete',
+    message_id: assistantMessageId,
+    chat_id: chatId,
+    tokens_used: updated?.tokens_used ?? null,
+    ctx_used: updated?.ctx_used ?? null,
+    ctx_max: updated?.ctx_max ?? null,
+    started_at: startedAt,
+    finished_at: updated?.finished_at ?? null,
+    model: session.model,
+  });
+
+  // Batch 9.7: ask_user_input pauses the loop. The tool row is still inserted
+  // (the answer endpoint needs a target row to UPDATE), but tool_results is
+  // pre-stamped with output=null as a "pending" sentinel and no tool_result
+  // frame goes out — the card renders from the tool_call frame alone. Mixed
+  // batches still execute the other tools normally.
+  ctx.publishUser({ type: 'chat_status', chat_id: chatId, status: 'tool_running', at: new Date().toISOString() });
+  let pausingForUserInput = false;
+  await Promise.all(
+    toolCalls.map(async (tc) => {
+      const [toolRow] = await ctx.sql<{ id: string }[]>`
+        INSERT INTO messages (session_id, chat_id, role, content, status, created_at)
+        VALUES (${sessionId}, ${chatId}, 'tool', '', 'complete', clock_timestamp())
+        RETURNING id
+      `;
+      const toolMessageId = toolRow!.id;
+      if (tc.name === 'ask_user_input') {
+        pausingForUserInput = true;
+        const sentinel = { tool_call_id: tc.id, output: null, truncated: false };
+        await ctx.sql`
+          UPDATE messages
+          SET tool_results = ${ctx.sql.json(sentinel as never)}
+          WHERE id = ${toolMessageId}
+        `;
+        // v1.13.0: mirror the pending sentinel into message_parts. The
+        // answer-endpoint UPDATE later (messages.ts:576) will delete and
+        // re-insert this part when the user submits their answer.
+        // TODO(v1.13.1): wrap the INSERT + UPDATE + insertParts triple in
+        // a per-iteration sql.begin before flipping read authority.
+        await insertParts(
+          ctx.sql,
+          partsFromToolMessage({ tool_results: sentinel }).map((p) => ({
+            ...p,
+            message_id: toolMessageId,
+          })),
+        );
+        return;
+      }
+      const tres = await executeToolCall(projectRoot, tc);
+      const stored = {
+        tool_call_id: tc.id,
+        output: tres.output,
+        truncated: tres.truncated,
+        ...(tres.error ? { error: tres.error } : {}),
+      };
+      await ctx.sql`
+        UPDATE messages
+        SET tool_results = ${ctx.sql.json(stored as never)}
+        WHERE id = ${toolMessageId}
+      `;
+      // v1.13.0: dual-write the tool_result part.
+      // TODO(v1.13.1): wrap the INSERT + UPDATE + insertParts triple in a
+      // per-iteration sql.begin before flipping read authority.
+      await insertParts(
+        ctx.sql,
+        partsFromToolMessage({ tool_results: stored }).map((p) => ({
+          ...p,
+          message_id: toolMessageId,
+        })),
+      );
+      ctx.publish(sessionId, {
+        type: 'tool_result',
+        tool_message_id: toolMessageId,
+        chat_id: chatId,
+        tool_call_id: tc.id,
+        output: tres.output,
+        truncated: tres.truncated,
+        ...(tres.error ? { error: tres.error } : {}),
+      });
+    })
+  );
+
+  if (pausingForUserInput) {
+    ctx.publishUser({
+      type: 'chat_status',
+      chat_id: chatId,
+      status: 'waiting_for_input',
+      at: new Date().toISOString(),
+    });
+    ctx.log.info(
+      { sessionId, chatId, assistantMessageId },
+      'inference paused awaiting user input',
+    );
+    return;
+  }
+
+  const [nextAssistant] = await ctx.sql<{ id: string }[]>`
+    INSERT INTO messages (session_id, chat_id, role, content, status, created_at)
+    VALUES (${sessionId}, ${chatId}, 'assistant', '', 'streaming', clock_timestamp())
+    RETURNING id
+  `;
+  await runAssistantTurn(ctx, {
+    sessionId,
+    chatId,
+    assistantMessageId: nextAssistant!.id,
+    // v1.8.2: charge this turn's actual tool invocations against the budget.
+    // One assistant message can emit multiple tool_calls, so we add the run
+    // count, not 1. The next turn's budget check sees the cumulative total.
+    toolsUsed: toolsUsed + result.toolCalls.length,
+    // v1.11.6: append the just-executed tool calls to the per-turn history
+    // so the next runAssistantTurn's doom-loop check can see them. We don't
+    // cap the array length here — per-turn budgets keep it bounded
+    // (typically <30 entries), and slicing happens inside detectDoomLoop.
+    recentToolCalls: [...args.recentToolCalls, ...result.toolCalls],
+    signal,
+  });
+}

apps/server/src/services/inference/turn.ts

@@ -0,0 +1,329 @@
+import type { FastifyBaseLogger } from 'fastify';
+import type { Sql } from '../../db.js';
+import type { Config } from '../../config.js';
+import type {
+  Agent,
+  ErrorReason,
+  Message,
+  MessageMetadata,
+  Project,
+  Session,
+  ToolCall,
+  UserStreamFrame,
+} from '../../types/api.js';
+import { ALL_TOOLS } from '../tools.js';
+import { resolveProjectRoot } from '../path_guard.js';
+import { maybeAutoNameChat } from '../auto_name.js';
+import { getAgentById } from '../agents.js';
+import * as compaction from '../compaction.js';
+import * as modelContext from '../model-context.js';
+import type { Broker } from '../broker.js';
+import { resolveToolBudget } from './budget.js';
+import {
+  DOOM_LOOP_THRESHOLD,
+  detectDoomLoop,
+} from './sentinels.js';
+import {
+  buildMessagesPayload,
+  loadContext,
+} from './payload.js';
+import {
+  finalizeCompletion,
+  handleAbortOrError,
+} from './error-handler.js';
+import {
+  executeStreamPhase,
+  streamCompletion,
+} from './stream-phase.js';
+import { executeToolPhase } from './tool-phase.js';
+import { DB_FLUSH_INTERVAL_MS, type StreamPhaseState } from './types.js';
+import {
+  runCapHitSummary,
+  runDoomLoopSummary,
+} from './sentinel-summaries.js';
+
+// v1.12.4: re-exported so external callers (tests, future consumers) keep
+// importing from services/inference.js as the public surface.
+export { detectDoomLoop, DOOM_LOOP_THRESHOLD } from './sentinels.js';
+export { buildMessagesPayload } from './payload.js';
+
+export interface InferenceFrame {
+  type:
+    | 'message_started'
+    | 'delta'
+    | 'tool_call'
+    | 'tool_result'
+    | 'message_complete'
+    | 'usage'
+    | 'messages_deleted'
+    | 'session_renamed'
+    | 'chat_renamed'
+    | 'error';
+  message_id?: string;
+  message_ids?: string[];
+  chat_id?: string;
+  tool_message_id?: string;
+  tool_call_id?: string;
+  // v1.8.2: 'system' added so cap-hit sentinel messages can announce themselves
+  // through the normal message_started → delta → message_complete sequence.
+  role?: 'assistant' | 'tool' | 'user' | 'system';
+  content?: string;
+  tool_call?: ToolCall;
+  output?: unknown;
+  truncated?: boolean;
+  error?: string;
+  // v1.8.2: structured error reason. Set on `type: 'error'` so the UI can
+  // surface a specific message; `error` stays the human-readable text.
+  reason?: ErrorReason;
+  // v1.8.2: piggybacks on `message_complete` so static or terminally-resolved
+  // messages can carry their persisted metadata to the live stream without a
+  // refetch (sentinels carry { kind: 'cap_hit', ... }; failed messages carry
+  // { kind: 'error', ... }).
+  metadata?: MessageMetadata | null;
+  tokens_used?: number | null;
+  ctx_used?: number | null;
+  ctx_max?: number | null;
+  completion_tokens?: number | null;
+  started_at?: string | null;
+  finished_at?: string | null;
+  model?: string;
+  session_id?: string;
+  name?: string;
+}
+
+export type FramePublisher = (sessionId: string, frame: InferenceFrame) => void;
+
+export interface InferenceContext {
+  sql: Sql;
+  config: Config;
+  log: FastifyBaseLogger;
+  publish: FramePublisher;
+  publishUser: (frame: UserStreamFrame) => void;
+  // v1.11: passed through so compaction.process can publish 'compacted'
+  // frames on the same session WS channel useSessionStream subscribes to.
+  // Compaction is the only path that needs the raw broker handle (regular
+  // inference goes through `publish`); keeping a separate field avoids
+  // tempting other code paths into bypassing the session-id binding.
+  broker: Broker;
+}
+
+// v1.12.4: payload assembly extracted to ./inference/payload.ts (tests
+// import buildMessagesPayload from this module, so a re-export below
+// preserves the public surface). Stream + tool phases extracted to
+// ./inference/stream-phase.ts and ./inference/tool-phase.ts.
+
+export interface StreamResult {
+  finishReason: string | null;
+  content: string;
+  toolCalls: ToolCall[];
+  promptTokens: number | null;
+  completionTokens: number | null;
+  // v1.13.1-C: reasoning text accumulated across reasoning-delta parts.
+  // Empty string when the model doesn't emit reasoning (most cases).
+  reasoning: string;
+}
+
+
+export interface TurnArgs {
+  sessionId: string;
+  chatId: string;
+  assistantMessageId: string;
+  // v1.8.2: cumulative tool calls executed this run. Compared against the
+  // resolved budget at the top of each turn. Replaces the older `depth`
+  // counter (which counted iterations, not invocations).
+  toolsUsed: number;
+  // v1.11.6: ordered tool calls executed in this user-message turn (across
+  // recursive runAssistantTurn invocations). Reset to [] at user-message
+  // boundaries by runInference, same as toolsUsed. Doom-loop check at the
+  // top of runAssistantTurn slices the last DOOM_LOOP_THRESHOLD entries.
+  recentToolCalls: ToolCall[];
+  signal: AbortSignal | undefined;
+}
+
+
+export async function runAssistantTurn(
+  ctx: InferenceContext,
+  args: TurnArgs,
+): Promise<void> {
+  const { sessionId, chatId } = args;
+
+  // v1.11: if the prior turn flagged this chat for compaction, run it first
+  // so loadContext below reads the post-compaction history. We swallow
+  // compaction failures (clearing the flag so we don't loop) and proceed
+  // with the un-compacted history — a slow turn that hits the model's
+  // hard limit is recoverable; a dead session is not.
+  const chatFlag = await ctx.sql<{ needs_compaction: boolean }[]>`
+    SELECT needs_compaction FROM chats WHERE id = ${chatId}
+  `;
+  if (chatFlag[0]?.needs_compaction) {
+    try {
+      await compaction.process({
+        sql: ctx.sql,
+        config: ctx.config,
+        log: ctx.log,
+        broker: ctx.broker,
+        chatId,
+      });
+    } catch (err) {
+      ctx.log.warn({ err, chatId }, 'auto-compaction failed; clearing flag and proceeding');
+      await ctx.sql`UPDATE chats SET needs_compaction = false WHERE id = ${chatId}`;
+    }
+  }
+
+  const loaded = await loadContext(ctx.sql, sessionId, chatId);
+  if (!loaded) {
+    ctx.log.warn({ sessionId }, 'inference: session or project missing');
+    return;
+  }
+  const { session, project, history } = loaded;
+  const projectRoot = await resolveProjectRoot(project.path);
+  // Agent resolution is per-turn so PATCH agent_id mid-conversation takes
+  // effect on the next message. Unknown agent_id returns null silently —
+  // session falls back to base prompt + all tools + default temperature.
+  const agent = session.agent_id
+    ? await getAgentById(project.path, session.agent_id)
+    : null;
+
+  // v1.8.2: cap-hit replaces the older "tool loop depth exceeded" failure.
+  // When we've already burned the budget *before* this turn even runs, we
+  // skip straight to the summary flow — the in-flight assistant message slot
+  // gets reused for the wrap-up reply instead of being marked failed.
+  const budget = resolveToolBudget(agent);
+  if (args.toolsUsed >= budget) {
+    await runCapHitSummary(ctx, args, session, project, history, agent, budget);
+    return;
+  }
+
+  // v1.11.6: doom-loop guard. Detected BEFORE the budget cap (the model can
+  // burn through 3 identical calls long before the 15-call budget fires).
+  // Same in-flight-slot-reuse pattern as runCapHitSummary — wrap-up reply
+  // lands in args.assistantMessageId, then a doom_loop sentinel is inserted
+  // to make the abort visible in the chat history.
+  const loop = detectDoomLoop(args.recentToolCalls);
+  if (loop) {
+    await runDoomLoopSummary(ctx, args, session, project, history, agent, loop);
+    return;
+  }
+
+  const messages = await buildMessagesPayload(session, project, history, agent);
+
+  // v1.11.8: resolve per-chat web-tools opt-in. Tri-state on the wire:
+  //   - session.web_search_enabled = null → inherit project default
+  //   - session.web_search_enabled = true/false → explicit
+  // Both web_search and web_fetch are gated by this single flag (the UI
+  // label is "Enable web search and fetch" — same store, both tools).
+  // Default is false unless explicitly opted in, matching the v1.9
+  // plumbing intent ("inert until Batch 8 ships the actual tools").
+  const webToolsEnabled =
+    session.web_search_enabled ?? project.default_web_search_enabled ?? false;
+
+  const state: StreamPhaseState = { accumulated: '', startedAt: null };
+  let result: StreamResult;
+  try {
+    result = await executeStreamPhase(ctx, args, session, messages, state, agent, webToolsEnabled);
+  } catch (err) {
+    await handleAbortOrError(ctx, args, state.accumulated, err);
+    return;
+  }
+
+  if (result.toolCalls.length > 0) {
+    await executeToolPhase(ctx, args, result, state.startedAt, session, projectRoot);
+    return;
+  }
+
+  await finalizeCompletion(ctx, args, result, state.startedAt, session);
+}
+
+export async function runInference(
+  ctx: InferenceContext,
+  sessionId: string,
+  chatId: string,
+  assistantMessageId: string,
+  signal?: AbortSignal
+): Promise<void> {
+  // v1.8.2: every fresh inference (initial send, regenerate, force_send,
+  // continue) starts with a clean budget. Tool-call accumulation across
+  // Continue invocations is what the hard ceiling guards against, not the
+  // per-call budget.
+  // v1.11.6: recentToolCalls also resets — doom-loop detection is scoped
+  // to a single user-message turn, so a Continue starts with no history.
+  return runAssistantTurn(ctx, {
+    sessionId,
+    chatId,
+    assistantMessageId,
+    toolsUsed: 0,
+    recentToolCalls: [],
+    signal,
+  });
+}
+
+// v1.8.2: cap-hit summary flow. Called instead of erroring when the loop
+// hits its budget. Reuses the in-flight assistant message slot to stream a
+// short wrap-up reply with the synthetic note prepended and tools disabled,
+// then always inserts a cap_hit sentinel afterward (regardless of summary
+// outcome) so the UI can show a Continue affordance.
+interface InferenceRegistration {
+  controller: AbortController;
+  completed: Promise<void>;
+}
+
+export function createInferenceRunner(
+  ctx: Omit<InferenceContext, 'publishUser'>,
+  publishUserFn: (user: string, frame: UserStreamFrame) => void
+) {
+  const registry = new Map<string, InferenceRegistration>();
+
+  return {
+    enqueue(sessionId: string, chatId: string, assistantMessageId: string, user: string) {
+      const callCtx: InferenceContext = {
+        ...ctx,
+        publishUser: (frame) => publishUserFn(user, frame),
+        // v1.11: broker comes in via ctx (set at registration time). Repeated
+        // here so the destructure carries it onto the per-call ctx without
+        // having to add it to every enqueue/cancel signature individually.
+        broker: ctx.broker,
+      };
+      // v1.8 mobile-tabs: announce working before the async loop starts so
+      // every device subscribed to the user channel sees the amber dot.
+      callCtx.publishUser({ type: 'chat_status', chat_id: chatId, status: 'streaming', at: new Date().toISOString() });
+      const controller = new AbortController();
+      let resolveCompleted!: () => void;
+      const completed = new Promise<void>((res) => { resolveCompleted = res; });
+      const registration: InferenceRegistration = { controller, completed };
+      registry.set(chatId, registration);
+      void (async () => {
+        try {
+          await runInference(callCtx, sessionId, chatId, assistantMessageId, controller.signal);
+          setImmediate(() => {
+            void maybeAutoNameChat(callCtx, chatId, sessionId).catch((err: Error) => {
+              callCtx.log.warn({ err, chatId }, 'auto-name failed');
+            });
+          });
+        } catch (err) {
+          callCtx.log.error({ err }, 'unhandled inference error');
+        } finally {
+          resolveCompleted();
+          // Only clear our own registration; a force-send may have replaced it.
+          if (registry.get(chatId) === registration) {
+            registry.delete(chatId);
+          }
+        }
+      })();
+    },
+
+    async cancel(_sessionId: string, chatId: string): Promise<boolean> {
+      const reg = registry.get(chatId);
+      if (!reg) return false;
+      reg.controller.abort();
+      // Swallow — we just need to wait for the catch/finally to persist state.
+      await reg.completed.catch(() => {});
+      return true;
+    },
+
+    hasActive(chatId: string): boolean {
+      return registry.has(chatId);
+    },
+  };
+}
+
+export const _toolNames = ALL_TOOLS.map((t) => t.name);

apps/server/src/services/inference/types.ts

@@ -0,0 +1,13 @@
+// v1.12.4: shared inter-phase types/constants for the extracted phase files.
+// Lives here so stream-phase, tool-phase, and the summary functions still in
+// inference.ts can all reference the same definitions without circular imports.
+
+export interface StreamPhaseState {
+  accumulated: string;
+  startedAt: string | null;
+}
+
+// 500ms keeps the DB UPDATE rate bounded under heavy streaming. Used by
+// executeStreamPhase, runCapHitSummary, and runDoomLoopSummary — every site
+// that does a debounced content flush during streaming.
+export const DB_FLUSH_INTERVAL_MS = 500;

apps/server/src/services/inference/xml-parser.ts

@@ -0,0 +1,53 @@
+// v1.10.5: XML-tag tool-call fallback. Some models emit
+// <tool_call><function=foo><parameter=key>value</parameter></function></tool_call>
+// in plain content instead of using the OpenAI tool_calls JSON channel.
+// The streaming loop in inference.ts extracts these blocks via these helpers.
+
+export const XML_TOOL_OPEN = '<tool_call>';
+export const XML_TOOL_CLOSE = '</tool_call>';
+
+export function parseXmlToolCall(
+  block: string,
+): { name: string; args: Record<string, unknown> } | null {
+  const nameMatch = block.match(/<function=([^>]+)>/);
+  if (!nameMatch || !nameMatch[1]) return null;
+  const name = nameMatch[1].trim();
+  if (!name) return null;
+  const args: Record<string, unknown> = {};
+  // Non-greedy body so each <parameter=…>…</parameter> pair is matched
+  // independently even when multiple appear in the same block.
+  const paramRe = /<parameter=([^>]+)>([\s\S]*?)<\/parameter>/g;
+  for (const m of block.matchAll(paramRe)) {
+    const key = (m[1] ?? '').trim();
+    if (!key) continue;
+    const raw = (m[2] ?? '').trim();
+    try {
+      args[key] = JSON.parse(raw);
+    } catch {
+      args[key] = raw;
+    }
+  }
+  return { name, args };
+}
+
+// Locate the first character that begins (or completely contains) an
+// unfinished <tool_call> opener in `s`. Returns -1 when `s` can be flushed
+// to the client in full without risking a partial tag leak.
+//   Case 1: a full `<tool_call>` opener with no matching closer — caller
+//           must keep everything from that index forward until the next
+//           chunk arrives with the closer.
+//   Case 2: `s` ends with a strict prefix of `<tool_call>` (e.g. `<tool_c`).
+//           Caller must keep just that suffix in the buffer.
+// Note: case 1 assumes the calling loop already extracted every complete
+// <tool_call>…</tool_call> pair before reaching this check.
+export function partialXmlOpenerStart(s: string): number {
+  const fullOpener = s.indexOf(XML_TOOL_OPEN);
+  if (fullOpener !== -1) return fullOpener;
+  const lastLt = s.lastIndexOf('<');
+  if (lastLt === -1) return -1;
+  const suffix = s.slice(lastLt);
+  if (XML_TOOL_OPEN.startsWith(suffix) && suffix.length < XML_TOOL_OPEN.length) {
+    return lastLt;
+  }
+  return -1;
+}

apps/server/src/services/tools.ts

@@ -527,6 +527,11 @@ export const askUserInput: ToolDef<AskUserInputInputT> = {
   },
 };
 
+// v1.13.3: alpha-sorted by tool.name at module load. llama.cpp's prompt
+// cache hits on byte-identical prefixes; the tool list lives near the top
+// of the system prompt, so any order drift would invalidate every cached
+// turn. Single source of truth for ordering lives here — toolJsonSchemas()
+// and TOOLS_BY_NAME inherit it.
 export const ALL_TOOLS: ReadonlyArray<ToolDef<unknown>> = [
   viewFile as ToolDef<unknown>,
   listDir as ToolDef<unknown>,
@@ -553,7 +558,7 @@ export const ALL_TOOLS: ReadonlyArray<ToolDef<unknown>> = [
   watchChanges as ToolDef<unknown>,
   getSemanticNeighborhoods as ToolDef<unknown>,
   getFrameworkAnalysis as ToolDef<unknown>,
-];
+].sort((a, b) => a.name.localeCompare(b.name));
 
 // v1.8.2: forward-compatible read-only whitelist. An agent whose `tools` is
 // fully contained in this set gets a generous default tool budget (30);

apps/server/src/types/api.ts

@@ -186,6 +186,11 @@ export interface Message {
   // v1.8.2: per-message metadata. See MessageMetadata for the discriminated
   // shapes currently in use.
   metadata: MessageMetadata | null;
+  // v1.13.1-C: reasoning content captured from the model's reasoning stream
+  // (qwen3.6 etc.). Populated from message_parts via the messages_with_parts
+  // view's reasoning_parts column. Optional — most rows have no reasoning
+  // and the API may omit the field on legacy responses.
+  reasoning_parts?: Array<{ text: string }> | null;
   // v1.11: anchored rolling compaction. Optional so consumers that SELECT
   // the pre-v1.11 column set still type-check. See compaction.ts +
   // schema.sql for semantics.

apps/web/src/api/client.ts

@@ -180,6 +180,11 @@ export const api = {
       request<{ ok: true }>(`/api/chats/${chatId}/compact`, { method: 'POST' }),
     stop: (chatId: string) =>
       request<{ stopped: boolean }>(`/api/chats/${chatId}/stop`, { method: 'POST' }),
+    discardStale: (chatId: string, messageId: string) =>
+      request<Message>(`/api/chats/${chatId}/discard_stale`, {
+        method: 'POST',
+        body: JSON.stringify({ message_id: messageId }),
+      }),
     forceSend: (chatId: string, content: string) =>
       request<{ user_message_id: string; assistant_message_id: string }>(
         `/api/chats/${chatId}/force_send`,

apps/web/src/api/types.ts

@@ -161,6 +161,11 @@ export interface Message {
   // v1.8.2: per-message metadata; see MessageMetadata. null for the vast
   // majority of messages.
   metadata: MessageMetadata | null;
+  // v1.13.1-C: reasoning content captured from models that stream reasoning
+  // tokens separately (qwen3.6 etc.). Backend populates from message_parts;
+  // optional on the wire — frontend doesn't render this yet (reserved for
+  // a v1.14 UI surface).
+  reasoning_parts?: Array<{ text: string }> | null;
   // v1.11: anchored rolling compaction fields. Optional on the wire so that
   // older API responses (or test fixtures) parse without explicit nulls.
   //   summary       — true on the assistant row that holds the active

apps/web/src/components/StaleStreamBanner.tsx

@@ -0,0 +1,34 @@
+interface Props {
+  onRetry: () => void;
+  onDiscard: () => void;
+}
+
+// v1.12.3: shown when an assistant message has been 'streaming' for 60+
+// seconds without new tokens. Lives above ChatInput in ChatPane. Retry
+// discards the stuck row then resends the last user message; Discard just
+// clears the row and drops the dot to idle.
+export function StaleStreamBanner({ onRetry, onDiscard }: Props) {
+  return (
+    <div className="border border-amber-500/30 bg-amber-500/5 rounded-md p-3 mb-2 mx-4 flex items-center justify-between gap-2">
+      <span className="text-sm text-muted-foreground">
+        Previous response didn't complete.
+      </span>
+      <div className="flex gap-2">
+        <button
+          type="button"
+          onClick={onRetry}
+          className="text-xs px-2 py-1 rounded border border-border hover:bg-accent max-md:min-h-[44px] max-md:px-3"
+        >
+          Retry
+        </button>
+        <button
+          type="button"
+          onClick={onDiscard}
+          className="text-xs px-2 py-1 rounded border border-border hover:bg-accent max-md:min-h-[44px] max-md:px-3"
+        >
+          Discard
+        </button>
+      </div>
+    </div>
+  );
+}

apps/web/src/components/panes/ChatPane.tsx

@@ -5,6 +5,7 @@ import { api } from '@/api/client';
 import { useSessionStream } from '@/hooks/useSessionStream';
 import { MessageList } from '@/components/MessageList';
 import { ChatInput } from '@/components/ChatInput';
+import { StaleStreamBanner } from '@/components/StaleStreamBanner';
 import {
   DropdownMenu,
   DropdownMenuContent,
@@ -44,6 +45,38 @@ export function ChatPane({ sessionId, chatId, projectId, agentId, onAgentChange,
 
   const chatMessages = stream.messages.filter((m) => m.chat_id === chatId);
   const streaming = chatMessages.some((m) => m.status === 'streaming');
+
+  // v1.12.3: stale-stream detection. Watches the (at most one) streaming
+  // assistant row. If its content length doesn't grow for STALE_THRESHOLD_MS,
+  // assume the upstream call is dead and surface the recovery banner. We use
+  // content length as the activity signal because every token delta extends
+  // it; last_seq isn't currently bumped per delta.
+  const STALE_THRESHOLD_MS = 60_000;
+  const streamingMsg = chatMessages.find((m) => m.status === 'streaming' && m.role === 'assistant');
+  const streamingId = streamingMsg?.id ?? null;
+  const streamingLen = streamingMsg?.content.length ?? 0;
+  const lastActivityRef = useRef<{ id: string; len: number; at: number } | null>(null);
+  const [stale, setStale] = useState(false);
+  useEffect(() => {
+    if (!streamingId) {
+      lastActivityRef.current = null;
+      setStale(false);
+      return;
+    }
+    const prev = lastActivityRef.current;
+    if (!prev || prev.id !== streamingId || prev.len !== streamingLen) {
+      lastActivityRef.current = { id: streamingId, len: streamingLen, at: Date.now() };
+      setStale(false);
+    }
+    const interval = setInterval(() => {
+      const a = lastActivityRef.current;
+      if (!a) return;
+      if (Date.now() - a.at >= STALE_THRESHOLD_MS) {
+        setStale(true);
+      }
+    }, 5_000);
+    return () => clearInterval(interval);
+  }, [streamingId, streamingLen]);
   // v1.11.5: per-chat model context limit comes from chat.model_context_limit
   // populated by GET /api/sessions/:id/chats. Threaded into ChatInput so
   // ContextBar can render a zero-state before the first assistant message.
@@ -87,6 +120,45 @@ export function ChatPane({ sessionId, chatId, projectId, agentId, onAgentChange,
     }
   }
 
+  const handleDiscardStale = useCallback(async () => {
+    if (!streamingId) return;
+    try {
+      await api.chats.discardStale(chatId, streamingId);
+      setStale(false);
+      lastActivityRef.current = null;
+    } catch (err) {
+      // 409 (race) is benign — the row already terminated some other way.
+      const msg = err instanceof Error ? err.message : 'discard failed';
+      if (!msg.includes('409')) toast.error(msg);
+      setStale(false);
+    }
+  }, [chatId, streamingId]);
+
+  const handleRetryStale = useCallback(async () => {
+    if (!streamingId) return;
+    const lastUser = [...chatMessages].reverse().find((m) => m.role === 'user' && m.kind === 'message');
+    if (!lastUser) {
+      toast.error('no prior user message to retry');
+      return;
+    }
+    try {
+      await api.chats.discardStale(chatId, streamingId);
+    } catch (err) {
+      const msg = err instanceof Error ? err.message : 'discard failed';
+      if (!msg.includes('409')) {
+        toast.error(msg);
+        return;
+      }
+    }
+    setStale(false);
+    lastActivityRef.current = null;
+    try {
+      await api.messages.send(chatId, lastUser.content);
+    } catch (err) {
+      toast.error(err instanceof Error ? err.message : 'retry send failed');
+    }
+  }, [chatId, streamingId, chatMessages]);
+
   const handleForceSend = useCallback(async (content: string) => {
     const trimmed = content.trim();
     if (!trimmed) return;
@@ -187,6 +259,13 @@ export function ChatPane({ sessionId, chatId, projectId, agentId, onAgentChange,
         </div>
       )}
 
+      {stale && streamingId && (
+        <StaleStreamBanner
+          onRetry={() => void handleRetryStale()}
+          onDiscard={() => void handleDiscardStale()}
+        />
+      )}
+
       <ChatInput
         disabled={false}
         projectId={projectId}

pnpm-lock.yaml

@@ -48,12 +48,18 @@ importers:
 
   apps/server:
     dependencies:
+      '@ai-sdk/openai-compatible':
+        specifier: ^2.0.47
+        version: 2.0.47(zod@3.25.76)
       '@fastify/static':
         specifier: ^7.0.4
         version: 7.0.4
       '@fastify/websocket':
         specifier: ^10.0.1
         version: 10.0.1
+      ai:
+        specifier: ^6.0.190
+        version: 6.0.190(zod@3.25.76)
       fastify:
         specifier: ^4.28.1
         version: 4.29.1
@@ -179,6 +185,28 @@ importers:
 
 packages:
 
+  '@ai-sdk/gateway@3.0.119':
+    resolution: {integrity: sha512-VAhfRWC+JexZakkVfmjaJKaTj00x7/UHdE8kMWL3NhuQAlf8oXtg9r4dfvFZrByXxchGRBvYE3biEUyibkg0xg==}
+    engines: {node: '>=18'}
+    peerDependencies:
+      zod: ^3.25.76 || ^4.1.8
+
+  '@ai-sdk/openai-compatible@2.0.47':
+    resolution: {integrity: sha512-Enm5UlL0zUCrW3792opk5h7hRWxZOZzDe6eQYVFqX9LUOGGCe1h8MZWAGim765nwzgnjlpeYOsuzZmLtRsTPlg==}
+    engines: {node: '>=18'}
+    peerDependencies:
+      zod: ^3.25.76 || ^4.1.8
+
+  '@ai-sdk/provider-utils@4.0.27':
+    resolution: {integrity: sha512-ubkAJ+xODouwtmN1tYlvTPphH1hPOBfZaEQe8U7skGvFAnIRs9PPpsq57bC2+Ky/MB4yzhd6YOsxTAx9sGpazw==}
+    engines: {node: '>=18'}
+    peerDependencies:
+      zod: ^3.25.76 || ^4.1.8
+
+  '@ai-sdk/provider@3.0.10':
+    resolution: {integrity: sha512-Q3BZ27qfpYqnCYGvE3vt+Qi6LGOF9R5Nmzn+9JoM1lCRsD9mYaIhfJLkSunN48nfGXJ6n+XNV0J/XVpqGQl7Dw==}
+    engines: {node: '>=18'}
+
   '@alloc/quick-lru@5.2.0':
     resolution: {integrity: sha512-UrcABB+4bUrFABwbluTIBErXwvbsU/V7TZWfmbgJfbkwiBuziS9gxdODUyuiecfdGQ85jglMW6juS3+z5TsKLw==}
     engines: {node: '>=10'}
@@ -789,6 +817,10 @@ packages:
   '@open-draft/until@2.1.0':
     resolution: {integrity: sha512-U69T3ItWHvLwGg5eJ0n3I62nWuE6ilHlmz7zM0npLBRvPRd7e6NYmg54vvRtP5mZG7kZqZCFVdsTWo7BPtBujg==}
 
+  '@opentelemetry/api@1.9.1':
+    resolution: {integrity: sha512-gLyJlPHPZYdAk1JENA9LeHejZe1Ti77/pTeFm/nMXmQH/HFZlcS/O2XJB+L8fkbrNSqhdtlvjBVjxwUYanNH5Q==}
+    engines: {node: '>=8.0.0'}
+
   '@pinojs/redact@0.4.0':
     resolution: {integrity: sha512-k2ENnmBugE/rzQfEcdWHcCY+/FM3VLzH9cYEsbdsoqrvzAKRhUZeRNhAZvB8OitQJ1TBed3yqWtdjzS6wJKBwg==}
 
@@ -1646,6 +1678,9 @@ packages:
     resolution: {integrity: sha512-tlqY9xq5ukxTUZBmoOp+m61cqwQD5pHJtFY3Mn8CA8ps6yghLH/Hw8UPdqg4OLmFW3IFlcXnQNmo/dh8HzXYIQ==}
     engines: {node: '>=18'}
 
+  '@standard-schema/spec@1.1.0':
+    resolution: {integrity: sha512-l2aFy5jALhniG5HgqrD6jXLi/rUWrKvqN/qJx6yoJsgKhblVd+iqqU4RCXavm/jPityDo5TCvKMnpjKnOriy0w==}
+
   '@tailwindcss/node@4.3.0':
     resolution: {integrity: sha512-aFb4gUhFOgdh9AXo4IzBEOzBkkAxm9VigwDJnMIYv3lcfXCJVesNfbEaBl4BNgVRyid92AmdviqwBUBRKSeY3g==}
 
@@ -1811,6 +1846,10 @@ packages:
   '@ungap/structured-clone@1.3.1':
     resolution: {integrity: sha512-mUFwbeTqrVgDQxFveS+df2yfap6iuP20NAKAsBt5jDEoOTDew+zwLAOilHCeQJOVSvmgCX4ogqIrA0mnyr08yQ==}
 
+  '@vercel/oidc@3.2.0':
+    resolution: {integrity: sha512-UycprH3T6n3jH0k44NHMa7pnFHGu/N05MjojYr+Mc6I7obkoLIJujSWwin1pCvdy/eOxrI/l3uDLQsmcrOb4ug==}
+    engines: {node: '>= 20'}
+
   '@vitejs/plugin-react@4.7.0':
     resolution: {integrity: sha512-gUu9hwfWvvEDBBmgtAowQCojwZmJ5mcLn3aufeCsitijs3+f2NsrPtlAWIR6OPiqljl96GVCUbLe0HyqIpVaoA==}
     engines: {node: ^14.18.0 || >=16.0.0}
@@ -1878,6 +1917,12 @@ packages:
     resolution: {integrity: sha512-MnA+YT8fwfJPgBx3m60MNqakm30XOkyIoH1y6huTQvC0PwZG7ki8NacLBcrPbNoo8vEZy7Jpuk7+jMO+CUovTQ==}
     engines: {node: '>= 14'}
 
+  ai@6.0.190:
+    resolution: {integrity: sha512-T+ixHbWZ6jmHRREpVVJTkFyWJeCekCdzLPan7lp1F32jG5OUw4+odlVYjtMRXVzogU+pWzpMmXdRiHUmdL/q0w==}
+    engines: {node: '>=18'}
+    peerDependencies:
+      zod: ^3.25.76 || ^4.1.8
+
   ajv-formats@2.1.1:
     resolution: {integrity: sha512-Wx0Kx52hxE7C18hkMEggYlEifqWZtYaRgouJor+WMdPnQyEK13vgEWyVNup7SoeeoLMsr4kf5h6dOW11I15MUA==}
     peerDependencies:
@@ -2694,6 +2739,9 @@ packages:
   json-schema-typed@8.0.2:
     resolution: {integrity: sha512-fQhoXdcvc3V28x7C7BMs4P5+kNlgUURe2jmUT1T//oBRMDrqy1QPelJimwZGo7Hg9VPV3EQV5Bnq4hbFy2vetA==}
 
+  json-schema@0.4.0:
+    resolution: {integrity: sha512-es94M3nTIfsEPisRafak+HDLfHXnKBhV3vU5eqPcS3flIWqcxJWgXHXiey3YrpaNsanY5ei1VoYEbOzijuq9BA==}
+
   json5@2.2.3:
     resolution: {integrity: sha512-XmOWe7eyHYH14cLdVPoyg+GOH3rYX++KpzrylJwSW98t3Nk+U8XOl8FWKOgwtzdb8lXGf6zYwDUzeHMWfxasyg==}
     engines: {node: '>=6'}
@@ -3966,6 +4014,30 @@ packages:
 
 snapshots:
 
+  '@ai-sdk/gateway@3.0.119(zod@3.25.76)':
+    dependencies:
+      '@ai-sdk/provider': 3.0.10
+      '@ai-sdk/provider-utils': 4.0.27(zod@3.25.76)
+      '@vercel/oidc': 3.2.0
+      zod: 3.25.76
+
+  '@ai-sdk/openai-compatible@2.0.47(zod@3.25.76)':
+    dependencies:
+      '@ai-sdk/provider': 3.0.10
+      '@ai-sdk/provider-utils': 4.0.27(zod@3.25.76)
+      zod: 3.25.76
+
+  '@ai-sdk/provider-utils@4.0.27(zod@3.25.76)':
+    dependencies:
+      '@ai-sdk/provider': 3.0.10
+      '@standard-schema/spec': 1.1.0
+      eventsource-parser: 3.0.8
+      zod: 3.25.76
+
+  '@ai-sdk/provider@3.0.10':
+    dependencies:
+      json-schema: 0.4.0
+
   '@alloc/quick-lru@5.2.0': {}
 
   '@babel/code-frame@7.29.0':
@@ -4516,6 +4588,8 @@ snapshots:
 
   '@open-draft/until@2.1.0': {}
 
+  '@opentelemetry/api@1.9.1': {}
+
   '@pinojs/redact@0.4.0': {}
 
   '@pkgjs/parseargs@0.11.0':
@@ -5386,6 +5460,8 @@ snapshots:
 
   '@sindresorhus/merge-streams@4.0.0': {}
 
+  '@standard-schema/spec@1.1.0': {}
+
   '@tailwindcss/node@4.3.0':
     dependencies:
       '@jridgewell/remapping': 2.3.5
@@ -5548,6 +5624,8 @@ snapshots:
 
   '@ungap/structured-clone@1.3.1': {}
 
+  '@vercel/oidc@3.2.0': {}
+
   '@vitejs/plugin-react@4.7.0(vite@5.4.21(@types/node@20.19.41)(lightningcss@1.32.0))':
     dependencies:
       '@babel/core': 7.29.0
@@ -5628,6 +5706,14 @@ snapshots:
 
   agent-base@7.1.4: {}
 
+  ai@6.0.190(zod@3.25.76):
+    dependencies:
+      '@ai-sdk/gateway': 3.0.119(zod@3.25.76)
+      '@ai-sdk/provider': 3.0.10
+      '@ai-sdk/provider-utils': 4.0.27(zod@3.25.76)
+      '@opentelemetry/api': 1.9.1
+      zod: 3.25.76
+
   ajv-formats@2.1.1(ajv@8.20.0):
     optionalDependencies:
       ajv: 8.20.0
@@ -6453,6 +6539,8 @@ snapshots:
 
   json-schema-typed@8.0.2: {}
 
+  json-schema@0.4.0: {}
+
   json5@2.2.3: {}
 
   jsonfile@6.2.1: