Add task control plane gates

[moose-lab]

Summary

• Add engine-neutral [DEVLOG_STAGE] and [DEVLOG_GATE] stdout parsing with persisted current_stage / gate_status state on tasks and sessions.
• Add human gate pause/resume handling through a dedicated resolve_gate session action, keeping approval replies isolated from the normal send queue.
• Surface stage and needs-input state on task cards, task details, session cards, and session headers, with global control-plane stream refresh for task/session lists.

Closes #37

Test Plan

• git diff --check origin/main..HEAD
• TZ=Asia/Shanghai bun run test src/core/__tests__/quality-gates.test.ts
• bun run typecheck
• bun run quality:ci
• Browser smoke on http://127.0.0.1:3000/tasks and /sessions: no console errors, no horizontal overflow

Audit Notes

• Verified T1.1-T4.2 from the Issue 让运行中的任务成为双向控制面：既能看到阶段进度，又能在需要人工确认的关卡介入 #37 work plan against current code and tests.
• Fixed one audit gap before PR: task/session lists now refresh from /api/devlog/stream control-plane events instead of relying only on polling.
• Left unrelated untracked docs/superpowers/plans/2026-06-03-session-agent-activity-stream.md out of the branch.

[claude]

Claude finished @moose-lab's task in 2m 47s —— View job

---

PR #38 Review — Add task control plane gates

Findings

• [must-fix] Duplicate parseGateStatus with diverging fallback — control-plane-state.ts (exported, used by UI) and control-plane-protocol.ts (private, used internally when reading gate_status back from DB) both implement parseGateStatus independently. They differ: the state module returns created_at: "" for a missing field; the protocol module returns new Date(0).toISOString(). Any future fix to one won't propagate to the other. The UI module should import from one canonical source. control-plane-state.ts:1-39 vs control-plane-protocol.ts:212-232. Fix this →

• [must-fix] migrateControlPlaneColumns called twice in getDb() — db.ts calls migrateControlPlaneColumns(_db) at line 40 (before the idle-status migration block) and again at line 179 (after the agent column migrations). On every cold start the PRAGMA table_info queries run twice and the columns-exist check runs twice per table. More importantly it's a correctness hazard: if any future migration between those two call sites modifies schema, the second call will see different state than expected. Fix this →

• [must-fix] resolveGate leaves isProcessing = true if writeGateResponse fails — process-manager.ts:1267 sets sp.isProcessing = true before calling writeGateResponse. writeGateResponse has an early-return error path (stdin not writable, line 1297–1302) that logs a warning but does not reset isProcessing. The session then blocks all new messages until the 3-minute watchdog fires. The gate_status has already been cleared from DB at this point, so the gate won't reappear. The fix is to reset sp.isProcessing = false in the early-return path of writeGateResponse, or reset it in the caller on write failure. Fix this →

• [should-fix] Gate resolve failure is silent in the UI — task-detail-dialog.tsx:209 returns early when !res.ok with no error state or toast shown to the user. If the gate resolve API returns a 409 (e.g., "no pending gate" — possible if the session was killed between render and submit), the button just stops spinning with no feedback. Add an error display. Fix this →

• [should-fix] Dual EventSource connections on pages that mount both hooks — use-tasks.ts and use-sessions.ts each independently open new EventSource("/api/devlog/stream"). On the Kanban board (/tasks), if both hooks are mounted simultaneously there are two SSE connections to the same endpoint. Each control-plane event triggers two independent refetch calls. This doubles backend load and causes brief double-render. A shared context or a single global SSE hook would fix this.

• [nit] Global status: running emit lacks session context — process-manager.ts:1290 emits streamManager.emit("global", { type: "status", status: "running" }). The status event shape doesn't include session_id, so a global subscriber can't tell which session went running. The per-session emit on line 1289 is fine; the global one is ambiguous. Compare to the control_plane_gate_resolved event on line 1280 which correctly includes session_id.

• [nit] isRecord helper defined three times — identical implementations in control-plane-protocol.ts:204, control-plane-state.ts:37, and process-manager.ts:515. Consider a shared @/core/utils export.

---

Overall assessment: The core architecture is clean — single-line marker parsing, safe JSON validation at boundaries, idempotent DB migrations, and proper SIGSTOP/SIGCONT gate pause are all well-done. Tests cover the protocol and state transitions thoroughly. The three must-fix items above are real defects (not just style); the rest are quality improvements.

Verdict: REQUEST_CHANGES