Correlate Codex turn dispatches with the turn the turn/start response names

[SawyerHood]

Human comments

What was wrong

Stacked on #3502 (merge that first). This layer removes the trigger behind #3462: the Codex bridge answered turn/start, then 250 ms later, if no Codex turn had opened yet, fabricated a completed zero-work turn (scheduleZeroWorkTurnSettlement). When the real turn/started arrived after that grace period, which happened on every one of the four production threads in #3462 (thread startup took ~40 s under load), the server saw a finished turn and the real work then ran as an unrequested root turn with no turn/input/accepted. That is #2580 (with #2634 folded in). The app-server's turn/start response already names the turn (TurnStartResponse { turn: { id, status } }), and the bridge discarded it.

What changed

This is Ratul Sarna's #2639, cherry-picked onto the #3502 branch with the protocol pin resolved to 205; the approach is unchanged and credited in the commit.

• plugins/provider-codex/src/bridge/bridge.ts: turn/start and thread/compact/start results are parsed; when the response names a turn, settleAcceptedDispatch opens that turn from the response (or only records input.accepted if the notification already opened it). The timer-based zero-work settlement now runs only for a turn/start result that names no turn.
• plugins/provider-codex/src/translator.ts: openTurnFromStartResponse emits the lifecycle from the response and dedupes the later turn/started/turn/completed notifications for that turn id; a response whose turn is already settled emits both boundaries.
• plugins/provider-codex/src/schemas.ts: exports codexTurnSchema.
• fake-codex-app-server.mjs: late start delay raised past the old grace period; new scripted prompts for respond-then-exit, respond-completed, steer-into-active, interrupt-before-start; compaction runs as a delayed turn, with env-selected modes for an idle report without a turn and an exit before the turn starts; a /late-start-interruptible prompt rejects an interrupt before turn/started like a real app-server.
• Compaction (added on top of Correlate codex turn dispatches with the turn the turn/start response names #2639): thread/compact/start answers {} and codex then opens the compaction as a turn (192 ms later in the committed real recording; a real app-server opens one even with nothing to compact and ends it failed). A compaction dispatch now waits for its turn/started and settles synthetically only on a signal from codex: thread/status/changed to idle (completed) or systemError (failed) while no turn is open, or the app-server exiting first (failed). No timer decides compaction.
• openTurnFromStartResponse no longer records a turn id in the already-open branch, where no later notification would remove it.
• Interrupts: opening a turn from the response exposes it before codex activates it. A real app-server answers turn/start with {turn: inProgress} and emits turn/started a few ms later; a turn/interrupt in that gap fails with no active turn to interrupt, so a stop right after bb saw the turn failed (the off-pipeline @bb/agent-runtime codex integration test "stops an active turn and recovers with a follow-up" hit it under load). The bridge now tracks response-opened turns and, only when an interrupt of one is rejected, waits for codex's own turn/started (or the turn settling or the app-server exiting, within the interrupt settlement timeout) and interrupts once more.
• HOST_DAEMON_PROTOCOL_VERSION 204 → 205 (on top of Reconcile daemon turns the server lost track of on explicit stop #3502's 204).

How you verified

• bridge.zero-work-turn.test.ts: the exactly-once late-start test now runs with a 350 ms delay and fails on main (two starts), plus new cases for a response followed by app-server exit, a response already completed, a steer into the active turn, and an interrupt before the notification.
• pnpm exec turbo run typecheck --filter=bb-plugin-provider-codex --filter=@bb/host-daemon-contract and turbo run test for the same packages: codex plugin 276/276 (27 files), contract 55/55, typecheck green. With bridge.ts, translator.ts and schemas.ts swapped back to the Reconcile daemon turns the server lost track of on explicit stop #3502 base, five of the seven zero-work tests fail (the late-start exactly-once case and the four new ones) and the two pre-existing ones pass.
• Live dev app with real Codex (codex CLI 0.154.0, device-auth login in an isolated scratch CODEX_HOME) behind a documented delay relay around the real codex app-server (an untracked codex shim holds turn/started and following notifications 400 ms while responses pass): on origin/main and on the Reconcile daemon turns the server lost track of on explicit stop #3502 base, every prompt produced two root turns within the same second, a synthetic one with turn/input/accepted and an instant completed, then the real turn with no accepted input (Codex threads 01a09199-7ab8…, 01a0919a-0f7e…). On this layer, under the same relay, three consecutive real turns (app-server ids 01a0919a-b048…, 01a0919b-7d59…, 01a0919b-b5b4…) each produced exactly one lifecycle with its accepted input, and a follow-up sent right after the first completion steered into the running turn instead of being fabricated or refused.
• Compaction regressions in bridge.zero-work-turn.test.ts: a turn/started 600 ms after the empty compact result keeps one lifecycle on the real turn; an idle status report without a turn settles it completed; an app-server exit before the turn settles it failed. With bridge.ts and translator.ts swapped back to the previous layer head, the 600 ms and exit cases fail (two lifecycles; a false completed), and the idle case passed there too (the old timer also completed it).
• On the rebased head: turbo run typecheck lint and turbo run test for bb-plugin-provider-codex (288), @bb/provider-bridge-protocol (238), @bb/provider-parity (56, including the Codex replay), @bb/host-daemon-contract (56); format check clean; EAP codename scan clean for the push range.
• Real Codex, delayed compaction (same scratch login, relay holding turn/started 800 ms): at the commit before the compaction fix, bb thread compact produced a synthetic turn accepted and completed in the same millisecond, then the real compaction as an unrequested second root turn with the contextCompaction item (real app-server turn 01a09269-fbad…). With the fix, the same request produced one root turn with the accepted input, the contextCompaction item and the real completion (real app-server turn 01a0926a-a9e6…).
• Interrupt regression in bridge.zero-work-turn.test.ts: a stop sent as soon as the response-opened turn is seen succeeds and settles interrupted; with bridge.ts swapped back to the previous layer head it fails with the -32000 no active turn to interrupt error.
• Real Codex, interrupt in the activation gap: probing the app-server directly, turn/interrupt sent right after the turn/start response returned -32600 no active turn to interrupt and the same call after turn/started succeeded. Driving the bridge against the real app-server and stopping each turn as soon as bb saw it started: on the previous head 1 of 3 stops failed with that error (Codex turn 01a092a5-9c02…); with the fix all 3 succeeded (01a092ac-269e…, 01a092ad-1391…, 01a092ae-0076…). The probe was a temporary test file and is not committed. With the fix, the @bb/agent-runtime codex integration file passes against real Codex (8 of 8, including the stop-and-recover and steer tests).

Fixes #2580
Closes #2639

🤖 Generated with Claude Code

AGENT GENERATED