[#975] De-block the event loop

[realproject7]

Closes #975

De-block the event loop. QuadWork runs as a single Node process, so any synchronous block on the main thread freezes every agent's WebSocket, HTTP, and timers at once. This removes the 3 confirmed stalls — behavior-preservingly (only the blocking is removed; outcomes are unchanged).

Changes

1. file-chat.js appendMessage busy-wait — the append retry loop slept via Atomics.wait(new Int32Array(new SharedArrayBuffer(4)), 0, 0, 100), a synchronous 100ms block on the main thread (up to 300ms across 3 attempts) on every chat + system message. Dropped the sleep — retries now run immediately, attempt count unchanged. appendFileSync transient failures are rare and not time-dependent, so the busy-wait bought nothing but latency.

2. index.js spawnButlerPty pre-trust — ran execFileSync("claude", ["-p", "echo ok"], { timeout: 15000 }) synchronously on the event loop; a hung claude blackholed the whole server for up to 15s. Dropped it: the interactive trust prompt is already auto-answered by the onData trust-listener installed a few lines below (the same mechanism the agent PTYs rely on), so no synchronous pre-trust is needed to reach a trusted session.

3. routes.js /api/setup route — the setup shell-outs (gh repo clone, git fetch, git worktree add, claude -p) ran synchronously in the request handler. ensureGitHeadForSetup and createAgentWorktree are now async, awaiting an injectable execFn (default the new execAsync, backed by the already-imported _execFileAsync = promisify(execFile)); the /api/setup handler is async and awaits them. Express 5 forwards a rejected async handler to the default 500 handler — the same result the prior synchronous throws produced. The now-dead synchronous exec() wrapper was removed (execFileSync is still used directly by the small gh api helpers).

EPIC Alignment (#967)

EPIC #967 Phase 3, the last fix ticket. Isolated to three de-block sites; the one change on the chat write path (file-chat) is behavior-preserving with ordering verified below. No agent-spawn / steady-state runtime change beyond removing blocking. JSONL rotation is explicitly out of scope (a separate future ticket, per the issue).

Self-Verification

• npm test (server suite, what CI runs) → 60 passed, 0 failed, 2 skipped. Both setup test files were migrated to inject an async fake execFn and await the now-async functions: routes.setupWorktrees.test.js 4/4 (stale/empty/clone HEAD paths), routes.setupHardening.test.js 6/6 (slug normalization, origin match/mismatch, worktree fresh/stale-reuse/detached).
• node --check on all 5 files → clean.
• Server boots + shuts down cleanly on this branch (temp HOME, free port) → prints listening, exits on SIGINT with no errors.
• Chat ordering + ids are unaffected (the ticket's key safety check): the message id is assigned synchronously (state.nextId++) before the append, and appendFileSync writes in call order — removing the inter-retry sleep changes neither. Verified at runtime: 50 sequential appendMessage calls → ids 1..50 contiguous and strictly monotonic, and the on-disk JSONL reads back in ascending id order. The existing file-chat.test.js / file-chat.system-messages.test.js / file-chat.loopguard.test.js all still pass.

🤖 Generated with Claude Code

[project7-interns]

Verdict: APPROVE

Epic Alignment: PASS

EPIC #967 Phase 3 calls for minimal local hardening without regressing the orchestrator; #975 specifically requires behavior-preserving removal of chat/butler/setup event-loop stalls, and the diff does exactly that.

Checked (evidence)

• Structural gate: PR body contains Closes #975, filled ## EPIC Alignment (#967), and filled ## Self-Verification.
• Context: gh issue view 975 and gh api repos/realproject7/quadwork/issues/967 read live; contract checked: behavior unchanged, only blocking removed, use async execFile path in routes.
• Chat path: server/file-chat.js:146 still assigns ids synchronously before append; server/file-chat.js:166 keeps 3 attempts; server/file-chat.js:178 removes the Atomics.wait sleep.
• Butler path: server/index.js:1009 removes the blocking pre-trust claude -p; server/index.js:1077 keeps the trust prompt listener.
• Setup path: server/routes.js:3691 uses promisified async execFile; server/routes.js:3711 and server/routes.js:3777 make setup helpers async; server/routes.js:3884 makes /api/setup async and awaits shell-outs at server/routes.js:3892, server/routes.js:3906, server/routes.js:3920, server/routes.js:3937.
• Tests updated for async setup helpers: server/routes.setupWorktrees.test.js and server/routes.setupHardening.test.js migrated to awaited fake execFns.
• Riskiest part of this diff: chat write path ordering; acceptable because id/seq assignment remains before append, append remains synchronous, and only inter-retry sleep was removed.
• Kill-list: scanned all items — clean.
• CI: gh pr checks 985 → test pass, 52s.

Findings

None.

Decision

Approved. The changes are scoped to the three #975 stalls, preserve the relevant contracts, and CI passes on current head 8d1b61dfb40863b899d2080178faf61eed3bab56.

[realproject7]

APPROVE at 8d1b61dfb40863b899d2080178faf61eed3bab56

Full independent review — issue #975 De-block the event loop (EPIC #967 Phase 3, last fix ticket). Behavior-preserving removal of 3 synchronous main-thread stalls; comment-based approval (shared reviewer GH identity blocks a formal Approve).

Checked (evidence)

• file-chat busy-wait removed (stall 1) — server/file-chat.js:178: Atomics.wait(...,100) dropped from the append retry loop; retry is now immediate, attempt count unchanged (retries = 3). Ordering/ids safe because id is assigned synchronously at :146 (state.nextId++) before any I/O and the whole appendMessage is synchronous (appendFileSync), so no interleaving.
• butler pre-trust removed (stall 2) — server/index.js:1006-1012: the blocking execFileSync("claude",["-p","echo ok"], timeout 15000) is gone; the interactive trust prompt is auto-answered by the onData trust-listener at server/index.js:1077-1091 (writes "1\r"), the same mechanism agent PTYs use. No sync pre-trust needed.
• setup shell-outs async (stall 3) — server/routes.js:3691 new execAsync (promisified _execFileAsync, imported :7) keeps the old {ok, output} contract; ensureGitHeadForSetup (:3711) and createAgentWorktree (:3777) are now async with injectable execFn = execAsync, all call sites await; handler router.post("/api/setup", async …) at :3884. Express 5.2.1 (package.json ^5.2.1) forwards a rejected async handler to the default error handler → same 500 as the prior sync throw. Removed sync exec() wrapper has no remaining callers.

Verification

• Chat ordering/ids (flagged invariant) — verified empirically: drove appendMessage 50× → returned ids contiguous+monotonic 1..50, on-disk JSONL ascending+contiguous 1..50, returned order === disk order.
• node --check clean on all 5 files; both migrated setup tests pass (setupHardening 6/6, setupWorktrees 4/4); full suite 60 passed, 0 failed (shared deps); CI test SUCCESS.

No findings.