@winspan/claude-forge 8.54.4 → 9.12.0

package/dist/daemon/services/spec-gate.d.ts

@@ -0,0 +1,54 @@
+/**
+ * SpecGate — orchestrates the approve-token short-circuit (step 0) inside
+ * `UserPromptHandler`.
+ *
+ * Extracted from `UserPromptHandler` (D-3 step 0). Stateless wrapper around
+ * `SpecApproval` so the handler can stay thin.
+ *
+ * The method is fail-silent unless documented otherwise — the prompt pipeline
+ * must keep running even if persistence hiccups.
+ *
+ * 2026-05-27 Spec 1702: `tryAutoGenerate` + `SpecGenerator` integration
+ * removed entirely (0% real-world acceptance rate → dead code).
+ * `SpecApproval` + `handleApproval` + `handleApproveToken` are preserved for
+ * the hand-written spec register path (planner agent calls `cf spec register`
+ * then user replies "批准").
+ */
+import type { UserPromptSubmitEvent } from '../../core/types.js';
+import type { SpecApproval } from './spec-approval.js';
+export declare class SpecGate {
+    private readonly specApproval;
+    constructor(specApproval?: SpecApproval | null);
+    /**
+     * Step 0 short-circuit: if the prompt is exactly `approve <8-hex>`,
+     * verify the token and return a system message. Returns `null` when the
+     * prompt is not an approve command (caller continues the regular pipeline).
+     *
+     * @deprecated Bug 1 (2026-05-27 spec § 3.1) — use {@link handleApproval}
+     *   instead, which also handles bare CJK keywords ('批准' / '同意') and
+     *   falls back to the latest pending row when no token is supplied. Kept
+     *   for backwards compatibility / unit tests that already cover the
+     *   token-only path.
+     */
+    handleApproveToken(event: UserPromptSubmitEvent, prompt: string): string | null;
+    /**
+     * Bug 1 fix (2026-05-27 spec § 3.1) — generalised approval handler.
+     *
+     * Recognises:
+     *   - 'approve <8-hex>'              → verify token directly
+     *   - '批准 <8-hex>' / '同意 <8-hex>' → verify token directly
+     *   - bare '批准' / '同意' / 'approve' → fallback to latest pending row in
+     *                                       the session, verify its token
+     *
+     * Returns:
+     *   - systemMessage string → caller short-circuits (skip KB / Skill / wf rec)
+     *   - null                  → not an approval signal; caller runs normal pipe
+     *
+     * Per spec Q6 (user-approved): even when fallback finds NO pending row,
+     * we STILL short-circuit with a [AUTO-SPEC] '本 session 没有 pending spec'
+     * systemMessage. Letting bare '批准' fall through to normal pipeline would
+     * eat the prompt as a no-op task title (Bug 1 root cause).
+     */
+    handleApproval(event: UserPromptSubmitEvent, prompt: string): string | null;
+}
+//# sourceMappingURL=spec-gate.d.ts.map

package/dist/daemon/services/spec-gate.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"spec-gate.d.ts","sourceRoot":"","sources":["../../../src/daemon/services/spec-gate.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,OAAO,KAAK,EAAE,qBAAqB,EAAE,MAAM,qBAAqB,CAAC;AACjE,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAIvD,qBAAa,QAAQ;IAEjB,OAAO,CAAC,QAAQ,CAAC,YAAY;gBAAZ,YAAY,GAAE,YAAY,GAAG,IAAW;IAG3D;;;;;;;;;;OAUG;IACH,kBAAkB,CAAC,KAAK,EAAE,qBAAqB,EAAE,MAAM,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI;IAkB/E;;;;;;;;;;;;;;;;;OAiBG;IACH,cAAc,CAAC,KAAK,EAAE,qBAAqB,EAAE,MAAM,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI;CAkD5E"}

package/dist/daemon/services/spec-gate.js

@@ -0,0 +1,113 @@
+/**
+ * SpecGate — orchestrates the approve-token short-circuit (step 0) inside
+ * `UserPromptHandler`.
+ *
+ * Extracted from `UserPromptHandler` (D-3 step 0). Stateless wrapper around
+ * `SpecApproval` so the handler can stay thin.
+ *
+ * The method is fail-silent unless documented otherwise — the prompt pipeline
+ * must keep running even if persistence hiccups.
+ *
+ * 2026-05-27 Spec 1702: `tryAutoGenerate` + `SpecGenerator` integration
+ * removed entirely (0% real-world acceptance rate → dead code).
+ * `SpecApproval` + `handleApproval` + `handleApproveToken` are preserved for
+ * the hand-written spec register path (planner agent calls `cf spec register`
+ * then user replies "批准").
+ */
+import { formatApprovalMessage } from './spec-approval.js';
+import { logger } from '../../core/utils/logger.js';
+export class SpecGate {
+    specApproval;
+    constructor(specApproval = null) {
+        this.specApproval = specApproval;
+    }
+    /**
+     * Step 0 short-circuit: if the prompt is exactly `approve <8-hex>`,
+     * verify the token and return a system message. Returns `null` when the
+     * prompt is not an approve command (caller continues the regular pipeline).
+     *
+     * @deprecated Bug 1 (2026-05-27 spec § 3.1) — use {@link handleApproval}
+     *   instead, which also handles bare CJK keywords ('批准' / '同意') and
+     *   falls back to the latest pending row when no token is supplied. Kept
+     *   for backwards compatibility / unit tests that already cover the
+     *   token-only path.
+     */
+    handleApproveToken(event, prompt) {
+        if (!this.specApproval)
+            return null;
+        const token = this.specApproval.extractToken(prompt);
+        if (!token)
+            return null;
+        const result = this.specApproval.verify(token, event.session_id, event.timestamp);
+        const pendingList = result.outcome === 'not_found'
+            ? this.specApproval.listSessionPending(event.session_id)
+            : undefined;
+        const msg = formatApprovalMessage(result.outcome, {
+            token,
+            row: result.row,
+            pendingList,
+        });
+        logger.info(`[SpecGate] Approve short-circuit: ${result.outcome} token=${token}`);
+        return msg;
+    }
+    /**
+     * Bug 1 fix (2026-05-27 spec § 3.1) — generalised approval handler.
+     *
+     * Recognises:
+     *   - 'approve <8-hex>'              → verify token directly
+     *   - '批准 <8-hex>' / '同意 <8-hex>' → verify token directly
+     *   - bare '批准' / '同意' / 'approve' → fallback to latest pending row in
+     *                                       the session, verify its token
+     *
+     * Returns:
+     *   - systemMessage string → caller short-circuits (skip KB / Skill / wf rec)
+     *   - null                  → not an approval signal; caller runs normal pipe
+     *
+     * Per spec Q6 (user-approved): even when fallback finds NO pending row,
+     * we STILL short-circuit with a [AUTO-SPEC] '本 session 没有 pending spec'
+     * systemMessage. Letting bare '批准' fall through to normal pipeline would
+     * eat the prompt as a no-op task title (Bug 1 root cause).
+     */
+    handleApproval(event, prompt) {
+        if (!this.specApproval)
+            return null;
+        const intent = this.specApproval.extractApprovalIntent(prompt);
+        if (!intent.hasIntent)
+            return null;
+        // Path A: caller supplied explicit token → direct verify (legacy behavior).
+        if (intent.token) {
+            const result = this.specApproval.verify(intent.token, event.session_id, event.timestamp);
+            const pendingList = result.outcome === 'not_found'
+                ? this.specApproval.listSessionPending(event.session_id)
+                : undefined;
+            const msg = formatApprovalMessage(result.outcome, {
+                token: intent.token,
+                row: result.row,
+                pendingList,
+            });
+            logger.info(`[SpecGate] Approval short-circuit (with-token): ${result.outcome} token=${intent.token}`);
+            return msg;
+        }
+        // Path B: bare keyword → fallback to the latest pending row for this
+        // session. listSessionPending already filters status='generated' DESC by
+        // created_at, so LIMIT 1 gives us the most recent unapproved spec.
+        const recent = this.specApproval.listSessionPending(event.session_id, 1);
+        if (recent.length === 0) {
+            // Spec Q6: no pending → return systemMessage saying so (DON'T fallthrough).
+            logger.info('[SpecGate] Approval short-circuit (no-token): no pending spec found');
+            return formatApprovalMessage('not_found', {
+                token: '(无 token)',
+                pendingList: [],
+            });
+        }
+        const candidate = recent[0];
+        const result = this.specApproval.verify(candidate.token, event.session_id, event.timestamp);
+        logger.info(`[SpecGate] Approval short-circuit (no-token fallback): ${result.outcome} ` +
+            `token=${candidate.token} spec=${candidate.spec_path}`);
+        return formatApprovalMessage(result.outcome, {
+            token: candidate.token,
+            row: result.row ?? candidate,
+        });
+    }
+}
+//# sourceMappingURL=spec-gate.js.map

package/dist/daemon/services/spec-gate.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"spec-gate.js","sourceRoot":"","sources":["../../../src/daemon/services/spec-gate.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAIH,OAAO,EAAE,qBAAqB,EAAE,MAAM,oBAAoB,CAAC;AAC3D,OAAO,EAAE,MAAM,EAAE,MAAM,4BAA4B,CAAC;AAEpD,MAAM,OAAO,QAAQ;IAEA;IADnB,YACmB,eAAoC,IAAI;QAAxC,iBAAY,GAAZ,YAAY,CAA4B;IACxD,CAAC;IAEJ;;;;;;;;;;OAUG;IACH,kBAAkB,CAAC,KAA4B,EAAE,MAAc;QAC7D,IAAI,CAAC,IAAI,CAAC,YAAY;YAAE,OAAO,IAAI,CAAC;QACpC,MAAM,KAAK,GAAG,IAAI,CAAC,YAAY,CAAC,YAAY,CAAC,MAAM,CAAC,CAAC;QACrD,IAAI,CAAC,KAAK;YAAE,OAAO,IAAI,CAAC;QAExB,MAAM,MAAM,GAAG,IAAI,CAAC,YAAY,CAAC,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,UAAU,EAAE,KAAK,CAAC,SAAS,CAAC,CAAC;QAClF,MAAM,WAAW,GAAG,MAAM,CAAC,OAAO,KAAK,WAAW;YAChD,CAAC,CAAC,IAAI,CAAC,YAAY,CAAC,kBAAkB,CAAC,KAAK,CAAC,UAAU,CAAC;YACxD,CAAC,CAAC,SAAS,CAAC;QACd,MAAM,GAAG,GAAG,qBAAqB,CAAC,MAAM,CAAC,OAAO,EAAE;YAChD,KAAK;YACL,GAAG,EAAE,MAAM,CAAC,GAAG;YACf,WAAW;SACZ,CAAC,CAAC;QACH,MAAM,CAAC,IAAI,CAAC,qCAAqC,MAAM,CAAC,OAAO,UAAU,KAAK,EAAE,CAAC,CAAC;QAClF,OAAO,GAAG,CAAC;IACb,CAAC;IAED;;;;;;;;;;;;;;;;;OAiBG;IACH,cAAc,CAAC,KAA4B,EAAE,MAAc;QACzD,IAAI,CAAC,IAAI,CAAC,YAAY;YAAE,OAAO,IAAI,CAAC;QACpC,MAAM,MAAM,GAAG,IAAI,CAAC,YAAY,CAAC,qBAAqB,CAAC,MAAM,CAAC,CAAC;QAC/D,IAAI,CAAC,MAAM,CAAC,SAAS;YAAE,OAAO,IAAI,CAAC;QAEnC,4EAA4E;QAC5E,IAAI,MAAM,CAAC,KAAK,EAAE,CAAC;YACjB,MAAM,MAAM,GAAG,IAAI,CAAC,YAAY,CAAC,MAAM,CAAC,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,UAAU,EAAE,KAAK,CAAC,SAAS,CAAC,CAAC;YACzF,MAAM,WAAW,GACf,MAAM,CAAC,OAAO,KAAK,WAAW;gBAC5B,CAAC,CAAC,IAAI,CAAC,YAAY,CAAC,kBAAkB,CAAC,KAAK,CAAC,UAAU,CAAC;gBACxD,CAAC,CAAC,SAAS,CAAC;YAChB,MAAM,GAAG,GAAG,qBAAqB,CAAC,MAAM,CAAC,OAAO,EAAE;gBAChD,KAAK,EAAE,MAAM,CAAC,KAAK;gBACnB,GAAG,EAAE,MAAM,CAAC,GAAG;gBACf,WAAW;aACZ,CAAC,CAAC;YACH,MAAM,CAAC,IAAI,CACT,mDAAmD,MAAM,CAAC,OAAO,UAAU,MAAM,CAAC,KAAK,EAAE,CAC1F,CAAC;YACF,OAAO,GAAG,CAAC;QACb,CAAC;QAED,qEAAqE;QACrE,yEAAyE;QACzE,mEAAmE;QACnE,MAAM,MAAM,GAAG,IAAI,CAAC,YAAY,CAAC,kBAAkB,CAAC,KAAK,CAAC,UAAU,EAAE,CAAC,CAAC,CAAC;QACzE,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YACxB,4EAA4E;YAC5E,MAAM,CAAC,IAAI,CAAC,qEAAqE,CAAC,CAAC;YACnF,OAAO,qBAAqB,CAAC,WAAW,EAAE;gBACxC,KAAK,EAAE,WAAW;gBAClB,WAAW,EAAE,EAAE;aAChB,CAAC,CAAC;QACL,CAAC;QACD,MAAM,SAAS,GAAG,MAAM,CAAC,CAAC,CAAE,CAAC;QAC7B,MAAM,MAAM,GAAG,IAAI,CAAC,YAAY,CAAC,MAAM,CACrC,SAAS,CAAC,KAAK,EACf,KAAK,CAAC,UAAU,EAChB,KAAK,CAAC,SAAS,CAChB,CAAC;QACF,MAAM,CAAC,IAAI,CACT,0DAA0D,MAAM,CAAC,OAAO,GAAG;YACzE,SAAS,SAAS,CAAC,KAAK,SAAS,SAAS,CAAC,SAAS,EAAE,CACzD,CAAC;QACF,OAAO,qBAAqB,CAAC,MAAM,CAAC,OAAO,EAAE;YAC3C,KAAK,EAAE,SAAS,CAAC,KAAK;YACtB,GAAG,EAAE,MAAM,CAAC,GAAG,IAAI,SAAS;SAC7B,CAAC,CAAC;IACL,CAAC;CACF"}

package/dist/daemon/services/task-boundary-classifier.d.ts

@@ -0,0 +1,78 @@
+/**
+ * TaskBoundaryClassifier — async LLM "new task vs continuation" re-attribution.
+ *
+ * decision 265f59d5 (Option A, 2026-06-08). Solves the residual case the
+ * synchronous lexical segmenter cannot: a SHORT, ambiguous follow-up prompt
+ * that is neither a clear ack ("继续"/"好的") nor carries a NEW_TASK_KEYWORD,
+ * e.g. "现在全链路是不是都顺了" vs "文件也看下" — lexically indistinguishable,
+ * yet one is a new task and the other a continuation.
+ *
+ * ── Hot-path contract (NEVER block UserPromptSubmit) ──────────────────────
+ * The synchronous segmenter ALWAYS runs first and gives the prompt a
+ * best-effort task (zero latency). THIS service runs entirely in the
+ * background, fire-and-forget. It spawns `claude` (haiku) only for the ~30%
+ * of prompts the segmenter flagged ambiguous, then LAZILY re-attributes:
+ *   • LLM says NEW_TASK  → move the prompt's event onto a brand-new task row
+ *     (only `task_events.task_id` FK is touched — no event delete, no content
+ *     mutation). See SQLiteStorage.reattributePromptToNewTask.
+ *   • LLM says CONTINUATION → leave the sync best-effort merge untouched, just
+ *     stamp `llm_classified_at` so a replay never re-spawns.
+ *
+ * Failure mode: ANY spawn / timeout / parse error → give up, keep the sync
+ * result, never throw. Mirrors OutcomeClassificationService's fire-and-forget
+ * discipline.
+ */
+import type { SQLiteStorage } from '../../core/storage/sqlite.js';
+/** Verdict the agent must return (exactly one of these tokens). */
+export type BoundaryVerdict = 'NEW_TASK' | 'CONTINUATION';
+/**
+ * Pluggable spawn function — production binds {@link spawnClaudeCli}; tests
+ * inject a stub that returns a canned verdict WITHOUT touching claude.
+ */
+export type BoundarySpawnFn = (prompt: string) => Promise<string>;
+/** A unit of work enqueued by the segmenter for the ambiguous prompt. */
+export interface BoundaryClassifyJob {
+    sessionId: string;
+    /** The task the sync path best-effort merged this prompt into. */
+    mergedTaskId: string;
+    /** The UserPromptSubmit event that should move if the verdict is NEW_TASK. */
+    eventId: string;
+    /** The ambiguous prompt text itself. */
+    prompt: string;
+    /** Event timestamp (ISO) — used as the new task's start_time. */
+    timestamp: string;
+}
+export declare class TaskBoundaryClassifier {
+    private readonly spawnFn;
+    /**
+     * @param spawnFn  optional spawn override (tests). Default spawns the
+     *                 `task-boundary-classifier` agent via claude haiku.
+     */
+    constructor(opts?: {
+        spawnFn?: BoundarySpawnFn;
+        timeoutMs?: number;
+        claudeBin?: string;
+    });
+    /**
+     * Classify one ambiguous prompt and, if NEW_TASK, re-attribute it.
+     * Fire-and-forget: never throws. Returns the new task id when it split,
+     * else null (CONTINUATION, idempotent skip, or any failure).
+     */
+    classify(job: BoundaryClassifyJob, storage: SQLiteStorage): Promise<string | null>;
+    private safePriorPrompts;
+}
+/**
+ * Build the bare classification prompt handed to the agent. The agent body
+ * (.claude/agents/task-boundary-classifier.md) instructs it to answer with a
+ * single token; we still keep the contract inline so a bare-prompt spawn
+ * (no agent template synced yet) also works.
+ */
+export declare function buildBoundaryPrompt(priorPrompts: string[], current: string): string;
+/**
+ * Extract the agent's single-token verdict from raw stdout. Tolerant: the
+ * model may wrap the token in punctuation / markdown / a sentence. We scan for
+ * the first NEW_TASK / CONTINUATION occurrence (case-insensitive). Returns null
+ * when neither token is present (caller keeps the sync merge).
+ */
+export declare function parseVerdict(raw: string): BoundaryVerdict | null;
+//# sourceMappingURL=task-boundary-classifier.d.ts.map

package/dist/daemon/services/task-boundary-classifier.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"task-boundary-classifier.d.ts","sourceRoot":"","sources":["../../../src/daemon/services/task-boundary-classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AAEH,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,8BAA8B,CAAC;AAOlE,mEAAmE;AACnE,MAAM,MAAM,eAAe,GAAG,UAAU,GAAG,cAAc,CAAC;AAE1D;;;GAGG;AACH,MAAM,MAAM,eAAe,GAAG,CAAC,MAAM,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,CAAC,CAAC;AAElE,yEAAyE;AACzE,MAAM,WAAW,mBAAmB;IAClC,SAAS,EAAE,MAAM,CAAC;IAClB,kEAAkE;IAClE,YAAY,EAAE,MAAM,CAAC;IACrB,8EAA8E;IAC9E,OAAO,EAAE,MAAM,CAAC;IAChB,wCAAwC;IACxC,MAAM,EAAE,MAAM,CAAC;IACf,iEAAiE;IACjE,SAAS,EAAE,MAAM,CAAC;CACnB;AAcD,qBAAa,sBAAsB;IACjC,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAkB;IAE1C;;;OAGG;gBACS,IAAI,GAAE;QAAE,OAAO,CAAC,EAAE,eAAe,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAA;KAAO;IAe5F;;;;OAIG;IACG,QAAQ,CAAC,GAAG,EAAE,mBAAmB,EAAE,OAAO,EAAE,aAAa,GAAG,OAAO,CAAC,MAAM,GAAG,IAAI,CAAC;IAgGxF,OAAO,CAAC,gBAAgB;CAQzB;AAID;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,YAAY,EAAE,MAAM,EAAE,EAAE,OAAO,EAAE,MAAM,GAAG,MAAM,CAcnF;AAED;;;;;GAKG;AACH,wBAAgB,YAAY,CAAC,GAAG,EAAE,MAAM,GAAG,eAAe,GAAG,IAAI,CAUhE"}

package/dist/daemon/services/task-boundary-classifier.js

@@ -0,0 +1,202 @@
+/**
+ * TaskBoundaryClassifier — async LLM "new task vs continuation" re-attribution.
+ *
+ * decision 265f59d5 (Option A, 2026-06-08). Solves the residual case the
+ * synchronous lexical segmenter cannot: a SHORT, ambiguous follow-up prompt
+ * that is neither a clear ack ("继续"/"好的") nor carries a NEW_TASK_KEYWORD,
+ * e.g. "现在全链路是不是都顺了" vs "文件也看下" — lexically indistinguishable,
+ * yet one is a new task and the other a continuation.
+ *
+ * ── Hot-path contract (NEVER block UserPromptSubmit) ──────────────────────
+ * The synchronous segmenter ALWAYS runs first and gives the prompt a
+ * best-effort task (zero latency). THIS service runs entirely in the
+ * background, fire-and-forget. It spawns `claude` (haiku) only for the ~30%
+ * of prompts the segmenter flagged ambiguous, then LAZILY re-attributes:
+ *   • LLM says NEW_TASK  → move the prompt's event onto a brand-new task row
+ *     (only `task_events.task_id` FK is touched — no event delete, no content
+ *     mutation). See SQLiteStorage.reattributePromptToNewTask.
+ *   • LLM says CONTINUATION → leave the sync best-effort merge untouched, just
+ *     stamp `llm_classified_at` so a replay never re-spawns.
+ *
+ * Failure mode: ANY spawn / timeout / parse error → give up, keep the sync
+ * result, never throw. Mirrors OutcomeClassificationService's fire-and-forget
+ * discipline.
+ */
+import { logger } from '../../core/utils/logger.js';
+import { formatError, truncateString } from '../../core/utils/format.js';
+import { truncateSessionId } from '../../core/utils/session.js';
+import { spawnClaudeCli } from '../../core/utils/claude-cli-spawn.js';
+import { randomUUID } from 'node:crypto';
+/**
+ * Spawn timeout (decision 5266724a, 2026-06-08). Bumped 8s → 30s: live `claude`
+ * classification spawns were measured at 6.8–26s, so an 8s cap timed out the
+ * majority of calls and silently degraded back to the sync lexical path. 30s
+ * comfortably covers the observed tail. Still fire-and-forget / background, so
+ * a slow spawn never blocks UserPromptSubmit.
+ */
+const DEFAULT_TIMEOUT_MS = 30_000;
+/** How many prior prompts of the merged task to hand the LLM as context. */
+const CONTEXT_PROMPT_LIMIT = 3;
+export class TaskBoundaryClassifier {
+    spawnFn;
+    /**
+     * @param spawnFn  optional spawn override (tests). Default spawns the
+     *                 `task-boundary-classifier` agent via claude haiku.
+     */
+    constructor(opts = {}) {
+        const timeoutMs = opts.timeoutMs ?? DEFAULT_TIMEOUT_MS;
+        this.spawnFn =
+            opts.spawnFn ??
+                ((prompt) => spawnClaudeCli({
+                    agent: 'task-boundary-classifier',
+                    prompt,
+                    timeoutMs,
+                    claudeBin: opts.claudeBin,
+                    // No tools — pure classification, fastest path. Allowed list empty
+                    // means the agent uses its own frontmatter (which declares none).
+                }));
+    }
+    /**
+     * Classify one ambiguous prompt and, if NEW_TASK, re-attribute it.
+     * Fire-and-forget: never throws. Returns the new task id when it split,
+     * else null (CONTINUATION, idempotent skip, or any failure).
+     */
+    async classify(job, storage) {
+        try {
+            // Idempotency short-circuit BEFORE spending a spawn: if the merged task
+            // is already stamped, a prior run (or replay) handled this prompt.
+            if (storage.isLlmClassified(job.mergedTaskId)) {
+                return null;
+            }
+            // INFO-level entry log so production can finally confirm the classifier
+            // is actually invoked. Before decision 083112fc EVERY branch below was
+            // logger.debug() (invisible at the daemon's default INFO level), so a
+            // classifier that never once spawned looked identical to one that ran
+            // fine — `llm_classified_at` stayed 0/1132 with zero log evidence.
+            logger.info(`[TaskBoundaryClassifier] classify called for session=${truncateSessionId(job.sessionId)} ` +
+                `mergedTask=${truncateString(job.mergedTaskId, 8)} prompt="${truncateString(job.prompt.replace(/\s+/g, ' ').trim(), 60)}"`);
+            const priorPrompts = this.safePriorPrompts(storage, job.mergedTaskId);
+            const agentPrompt = buildBoundaryPrompt(priorPrompts, job.prompt);
+            let raw;
+            try {
+                raw = await this.spawnFn(agentPrompt);
+            }
+            catch (err) {
+                // NEVER swallow spawn failures silently. This is the single most
+                // important observability fix: a launchd PATH miss / exit error /
+                // timeout used to vanish at debug level, so "the classifier never
+                // ran" had no diagnosable cause. warn carries the error reason.
+                logger.warn(`[TaskBoundaryClassifier] spawn FAILED for session=${truncateSessionId(job.sessionId)}: ${formatError(err)} — keeping sync merge`);
+                return null; // conservative: keep sync best-effort merge
+            }
+            const verdict = parseVerdict(raw);
+            if (verdict === null) {
+                logger.warn(`[TaskBoundaryClassifier] unparseable verdict (raw="${truncateString(raw.trim(), 60)}") for ` +
+                    `session=${truncateSessionId(job.sessionId)} — keeping merge`);
+                return null;
+            }
+            logger.info(`[TaskBoundaryClassifier] verdict=${verdict} for session=${truncateSessionId(job.sessionId)} ` +
+                `mergedTask=${truncateString(job.mergedTaskId, 8)}`);
+            const atIso = new Date().toISOString();
+            if (verdict === 'CONTINUATION') {
+                // Correct already — just stamp so we never reprocess this prompt.
+                try {
+                    storage.markLlmClassified(job.mergedTaskId, atIso);
+                }
+                catch (err) {
+                    logger.warn(`[TaskBoundaryClassifier] stamp failed: ${formatError(err)}`);
+                }
+                return null;
+            }
+            // NEW_TASK → lazily re-attribute. reattributePromptToNewTask re-checks
+            // idempotency + the event's current FK inside one transaction (OQ #3).
+            const newTaskId = storage.reattributePromptToNewTask({
+                fromTaskId: job.mergedTaskId,
+                eventId: job.eventId,
+                newTask: {
+                    id: randomUUID(),
+                    session_id: job.sessionId,
+                    title: extractTitle(job.prompt),
+                    start_time: job.timestamp,
+                    task_kind: 'user',
+                },
+                atIso,
+            });
+            if (newTaskId) {
+                logger.info(`[TaskBoundaryClassifier] split ambiguous prompt off ${truncateString(job.mergedTaskId, 8)} ` +
+                    `→ new task ${truncateString(newTaskId, 8)} (session=${truncateSessionId(job.sessionId)})`);
+            }
+            else {
+                // reattribute can no-op if the FK moved under us (concurrent replay) —
+                // surface it so a silent "NEW_TASK verdict but no split" is diagnosable.
+                logger.info(`[TaskBoundaryClassifier] NEW_TASK verdict but reattribute no-op (idempotent skip) for ` +
+                    `session=${truncateSessionId(job.sessionId)}`);
+            }
+            return newTaskId;
+        }
+        catch (err) {
+            logger.warn(`[TaskBoundaryClassifier] classify failed: ${formatError(err)}`);
+            return null;
+        }
+    }
+    safePriorPrompts(storage, taskId) {
+        try {
+            return storage.getTaskUserPrompts(taskId, CONTEXT_PROMPT_LIMIT);
+        }
+        catch (err) {
+            logger.warn(`[TaskBoundaryClassifier] prior-prompt fetch failed: ${formatError(err)}`);
+            return [];
+        }
+    }
+}
+// ── Pure helpers (exported for unit-test reuse) ─────────────────────────────
+/**
+ * Build the bare classification prompt handed to the agent. The agent body
+ * (.claude/agents/task-boundary-classifier.md) instructs it to answer with a
+ * single token; we still keep the contract inline so a bare-prompt spawn
+ * (no agent template synced yet) also works.
+ */
+export function buildBoundaryPrompt(priorPrompts, current) {
+    const prev = priorPrompts.length > 0
+        ? priorPrompts.map((p, i) => `  ${i + 1}. ${truncateString(p.replace(/\s+/g, ' ').trim(), 200)}`).join('\n')
+        : '  (无)';
+    return [
+        '你是 task-boundary 分类器。判断「当前 prompt」是开启一个新任务，还是对上一个任务的追问/续接。',
+        '只输出一个词：NEW_TASK 或 CONTINUATION。不要解释，不要输出其他任何内容。',
+        '',
+        '上一个任务的若干 prompt：',
+        prev,
+        '',
+        `当前 prompt：${current.replace(/\s+/g, ' ').trim()}`,
+    ].join('\n');
+}
+/**
+ * Extract the agent's single-token verdict from raw stdout. Tolerant: the
+ * model may wrap the token in punctuation / markdown / a sentence. We scan for
+ * the first NEW_TASK / CONTINUATION occurrence (case-insensitive). Returns null
+ * when neither token is present (caller keeps the sync merge).
+ */
+export function parseVerdict(raw) {
+    if (!raw)
+        return null;
+    const upper = raw.toUpperCase();
+    const newIdx = upper.indexOf('NEW_TASK');
+    const contIdx = upper.indexOf('CONTINUATION');
+    if (newIdx < 0 && contIdx < 0)
+        return null;
+    if (newIdx < 0)
+        return 'CONTINUATION';
+    if (contIdx < 0)
+        return 'NEW_TASK';
+    // Both present — first one wins.
+    return newIdx < contIdx ? 'NEW_TASK' : 'CONTINUATION';
+}
+/** Title for a re-attributed prompt's new task (mirrors TaskSegmenter). */
+function extractTitle(prompt) {
+    const cleaned = prompt.replace(/\[Image #\d+\]/g, '').trim();
+    const firstLine = cleaned.split('\n')[0].trim();
+    if (firstLine.length <= 60)
+        return firstLine || '(无标题)';
+    return truncateString(firstLine, 60);
+}
+//# sourceMappingURL=task-boundary-classifier.js.map

package/dist/daemon/services/task-boundary-classifier.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"task-boundary-classifier.js","sourceRoot":"","sources":["../../../src/daemon/services/task-boundary-classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AAGH,OAAO,EAAE,MAAM,EAAE,MAAM,4BAA4B,CAAC;AACpD,OAAO,EAAE,WAAW,EAAE,cAAc,EAAE,MAAM,4BAA4B,CAAC;AACzE,OAAO,EAAE,iBAAiB,EAAE,MAAM,6BAA6B,CAAC;AAChE,OAAO,EAAE,cAAc,EAAE,MAAM,sCAAsC,CAAC;AACtE,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAwBzC;;;;;;GAMG;AACH,MAAM,kBAAkB,GAAG,MAAM,CAAC;AAElC,4EAA4E;AAC5E,MAAM,oBAAoB,GAAG,CAAC,CAAC;AAE/B,MAAM,OAAO,sBAAsB;IAChB,OAAO,CAAkB;IAE1C;;;OAGG;IACH,YAAY,OAA8E,EAAE;QAC1F,MAAM,SAAS,GAAG,IAAI,CAAC,SAAS,IAAI,kBAAkB,CAAC;QACvD,IAAI,CAAC,OAAO;YACV,IAAI,CAAC,OAAO;gBACZ,CAAC,CAAC,MAAc,EAAE,EAAE,CAClB,cAAc,CAAC;oBACb,KAAK,EAAE,0BAA0B;oBACjC,MAAM;oBACN,SAAS;oBACT,SAAS,EAAE,IAAI,CAAC,SAAS;oBACzB,mEAAmE;oBACnE,kEAAkE;iBACnE,CAAC,CAAC,CAAC;IACV,CAAC;IAED;;;;OAIG;IACH,KAAK,CAAC,QAAQ,CAAC,GAAwB,EAAE,OAAsB;QAC7D,IAAI,CAAC;YACH,wEAAwE;YACxE,mEAAmE;YACnE,IAAI,OAAO,CAAC,eAAe,CAAC,GAAG,CAAC,YAAY,CAAC,EAAE,CAAC;gBAC9C,OAAO,IAAI,CAAC;YACd,CAAC;YAED,wEAAwE;YACxE,uEAAuE;YACvE,sEAAsE;YACtE,sEAAsE;YACtE,mEAAmE;YACnE,MAAM,CAAC,IAAI,CACT,wDAAwD,iBAAiB,CAAC,GAAG,CAAC,SAAS,CAAC,GAAG;gBACzF,cAAc,cAAc,CAAC,GAAG,CAAC,YAAY,EAAE,CAAC,CAAC,YAAY,cAAc,CAAC,GAAG,CAAC,MAAM,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,EAAE,CAAC,GAAG,CAC7H,CAAC;YAEF,MAAM,YAAY,GAAG,IAAI,CAAC,gBAAgB,CAAC,OAAO,EAAE,GAAG,CAAC,YAAY,CAAC,CAAC;YACtE,MAAM,WAAW,GAAG,mBAAmB,CAAC,YAAY,EAAE,GAAG,CAAC,MAAM,CAAC,CAAC;YAElE,IAAI,GAAW,CAAC;YAChB,IAAI,CAAC;gBACH,GAAG,GAAG,MAAM,IAAI,CAAC,OAAO,CAAC,WAAW,CAAC,CAAC;YACxC,CAAC;YAAC,OAAO,GAAG,EAAE,CAAC;gBACb,iEAAiE;gBACjE,kEAAkE;gBAClE,kEAAkE;gBAClE,gEAAgE;gBAChE,MAAM,CAAC,IAAI,CACT,qDAAqD,iBAAiB,CAAC,GAAG,CAAC,SAAS,CAAC,KAAK,WAAW,CAAC,GAAG,CAAC,uBAAuB,CAClI,CAAC;gBACF,OAAO,IAAI,CAAC,CAAC,4CAA4C;YAC3D,CAAC;YAED,MAAM,OAAO,GAAG,YAAY,CAAC,GAAG,CAAC,CAAC;YAClC,IAAI,OAAO,KAAK,IAAI,EAAE,CAAC;gBACrB,MAAM,CAAC,IAAI,CACT,sDAAsD,cAAc,CAAC,GAAG,CAAC,IAAI,EAAE,EAAE,EAAE,CAAC,SAAS;oBAC3F,WAAW,iBAAiB,CAAC,GAAG,CAAC,SAAS,CAAC,kBAAkB,CAChE,CAAC;gBACF,OAAO,IAAI,CAAC;YACd,CAAC;YAED,MAAM,CAAC,IAAI,CACT,oCAAoC,OAAO,gBAAgB,iBAAiB,CAAC,GAAG,CAAC,SAAS,CAAC,GAAG;gBAC5F,cAAc,cAAc,CAAC,GAAG,CAAC,YAAY,EAAE,CAAC,CAAC,EAAE,CACtD,CAAC;YAEF,MAAM,KAAK,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;YAEvC,IAAI,OAAO,KAAK,cAAc,EAAE,CAAC;gBAC/B,kEAAkE;gBAClE,IAAI,CAAC;oBACH,OAAO,CAAC,iBAAiB,CAAC,GAAG,CAAC,YAAY,EAAE,KAAK,CAAC,CAAC;gBACrD,CAAC;gBAAC,OAAO,GAAG,EAAE,CAAC;oBACb,MAAM,CAAC,IAAI,CAAC,0CAA0C,WAAW,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;gBAC5E,CAAC;gBACD,OAAO,IAAI,CAAC;YACd,CAAC;YAED,uEAAuE;YACvE,uEAAuE;YACvE,MAAM,SAAS,GAAG,OAAO,CAAC,0BAA0B,CAAC;gBACnD,UAAU,EAAE,GAAG,CAAC,YAAY;gBAC5B,OAAO,EAAE,GAAG,CAAC,OAAO;gBACpB,OAAO,EAAE;oBACP,EAAE,EAAE,UAAU,EAAE;oBAChB,UAAU,EAAE,GAAG,CAAC,SAAS;oBACzB,KAAK,EAAE,YAAY,CAAC,GAAG,CAAC,MAAM,CAAC;oBAC/B,UAAU,EAAE,GAAG,CAAC,SAAS;oBACzB,SAAS,EAAE,MAAM;iBAClB;gBACD,KAAK;aACN,CAAC,CAAC;YAEH,IAAI,SAAS,EAAE,CAAC;gBACd,MAAM,CAAC,IAAI,CACT,uDAAuD,cAAc,CAAC,GAAG,CAAC,YAAY,EAAE,CAAC,CAAC,GAAG;oBAC3F,cAAc,cAAc,CAAC,SAAS,EAAE,CAAC,CAAC,aAAa,iBAAiB,CAAC,GAAG,CAAC,SAAS,CAAC,GAAG,CAC7F,CAAC;YACJ,CAAC;iBAAM,CAAC;gBACN,uEAAuE;gBACvE,yEAAyE;gBACzE,MAAM,CAAC,IAAI,CACT,wFAAwF;oBACtF,WAAW,iBAAiB,CAAC,GAAG,CAAC,SAAS,CAAC,EAAE,CAChD,CAAC;YACJ,CAAC;YACD,OAAO,SAAS,CAAC;QACnB,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,CAAC,IAAI,CAAC,6CAA6C,WAAW,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YAC7E,OAAO,IAAI,CAAC;QACd,CAAC;IACH,CAAC;IAEO,gBAAgB,CAAC,OAAsB,EAAE,MAAc;QAC7D,IAAI,CAAC;YACH,OAAO,OAAO,CAAC,kBAAkB,CAAC,MAAM,EAAE,oBAAoB,CAAC,CAAC;QAClE,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,MAAM,CAAC,IAAI,CAAC,uDAAuD,WAAW,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;YACvF,OAAO,EAAE,CAAC;QACZ,CAAC;IACH,CAAC;CACF;AAED,+EAA+E;AAE/E;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,YAAsB,EAAE,OAAe;IACzE,MAAM,IAAI,GACR,YAAY,CAAC,MAAM,GAAG,CAAC;QACrB,CAAC,CAAC,YAAY,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,KAAK,CAAC,GAAG,CAAC,KAAK,cAAc,CAAC,CAAC,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;QAC5G,CAAC,CAAC,OAAO,CAAC;IACd,OAAO;QACL,4DAA4D;QAC5D,iDAAiD;QACjD,EAAE;QACF,kBAAkB;QAClB,IAAI;QACJ,EAAE;QACF,aAAa,OAAO,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE;KACnD,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AACf,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,YAAY,CAAC,GAAW;IACtC,IAAI,CAAC,GAAG;QAAE,OAAO,IAAI,CAAC;IACtB,MAAM,KAAK,GAAG,GAAG,CAAC,WAAW,EAAE,CAAC;IAChC,MAAM,MAAM,GAAG,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC;IACzC,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO,CAAC,cAAc,CAAC,CAAC;IAC9C,IAAI,MAAM,GAAG,CAAC,IAAI,OAAO,GAAG,CAAC;QAAE,OAAO,IAAI,CAAC;IAC3C,IAAI,MAAM,GAAG,CAAC;QAAE,OAAO,cAAc,CAAC;IACtC,IAAI,OAAO,GAAG,CAAC;QAAE,OAAO,UAAU,CAAC;IACnC,iCAAiC;IACjC,OAAO,MAAM,GAAG,OAAO,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,cAAc,CAAC;AACxD,CAAC;AAED,2EAA2E;AAC3E,SAAS,YAAY,CAAC,MAAc;IAClC,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC,iBAAiB,EAAE,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC;IAC7D,MAAM,SAAS,GAAG,OAAO,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;IAChD,IAAI,SAAS,CAAC,MAAM,IAAI,EAAE;QAAE,OAAO,SAAS,IAAI,OAAO,CAAC;IACxD,OAAO,cAAc,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC;AACvC,CAAC"}