[LinuxDevil]

What changed

A client that loses its socket (or reloads) mid-turn reconnects, sends chat.resume, gets the missed events of a running turn, and otherwise reconciles from the stored transcript. Settled turns are never replayed.

• shared/authorWire.ts: chat.resume gains an optional afterSeq cursor; new AuthorSessionRecord extends AuthorSession { runningRequestId } is what GET /assistant/sessions/:id now returns.
• server/author/replayBuffer.ts (new): per-session ring of the last turn's AuthorServerEvents. turn.started resets the entry; overflow past REPLAY_CAPACITY (4000) evicts the oldest text.deltas first so every tool.* survives; a settled entry answers only its terminal event and is evicted lazily after REPLAY_SETTLED_TTL_MS (120 s).
• server/author/authorStream.ts: one ReplayBuffer per server records every runner event before broadcast; chat.resume is answered on the asking socket only — replay sends the buffered events with seq > afterSeq (live events follow on the same socket), settled sends only the terminal event, unknown sends an UNKNOWN_REQUEST turn.error (seq 0). Malformed afterSeq is a BAD_MESSAGE.
• server/author/routes.ts: GET /assistant/sessions/:id returns AuthorSessionRecord with runningRequestId from runner.running(id); runtimeClient.getAuthorSession retyped accordingly.
• src/author/reattach.ts (new): the pure helpers the chat hook will drive — streamEndRecovery (stoppedByUser → none, BUSY → busy, drop mid-stream → drop), acceptEvent seq dedupe (seq 0 always applies; new requestId always applies; same requestId only past lastSeq), pending-request localStorage helpers with try/catch around every access, and the reattach timeout/retry/gave-up constants.
• Removed four stale fallow-ignore-next-line unused-export suppressions in src/author/authorStream.ts that fallow reports as stale (the exports are used by the existing test entry point).

How it was tested

• server/author/__tests__/replayBuffer.test.ts (new, 7 tests): replay-after-cursor, unknown requestId/session/replaced turn, settled → terminal only (both turn.finished and turn.error), settled TTL expiry at exactly 120 s + 1 (fake clock via the now seam), running turns don't expire, and a 5 101-event overflow keeps all 2 500 tool.* events while dropping the oldest 1 101 text.deltas.
• server/__tests__/authorStream.test.ts (extended, 8 tests): a gated provider parks a turn mid-stream; the client drops after seq 3, reconnects, and chat.resume { afterSeq: 3 } delivers seqs 4–5 from the buffer then 6–7 live with no duplicates; a resume after settle delivers only the terminal event; an unknown requestId gets UNKNOWN_REQUEST. Also asserts GET /assistant/sessions/:id reports runningRequestId while the turn runs and null after.
• src/author/__tests__/reattach.test.ts (new, 12 tests): the streamEndRecovery table, acceptEvent dedupe cases, storage round-trip/malformed data, and a throwing localStorage.
• server/__tests__/authorSessions.test.ts: updated the GET /:id body assertion for the new runningRequestId field.

Command results:

• Studio npm run typecheck ✓ · npm run typecheck:server ✓ · npm test -- --run ✓ 968 passed · npm run test:server ✓ 1012 passed · npm run lint ✓ (0 errors)
• Root npx tsc --noEmit ✓ · npm run lint ✓ · npm run test:coverage ✓ 7163 passed · npm run fallow ✓ · npm run docs:ratchet ✓ · npm run api:check ✓ · npm run docs:llms:check ✓ · npm run size-check ✓

Acceptance criteria

• Reconnect mid-turn + chat.resume { afterSeq: 3 } delivers events 4…n then the live tail, with no duplicates.
• Resume after the turn settled sends only the terminal event (no text.delta, no tool.*).
• Resume with an unknown requestId → UNKNOWN_REQUEST.
• A settled entry older than 120 s answers unknown.
• Capacity overflow drops text.delta first and keeps every tool.* event.

Generated with Devin