M7: Skip the redundant Codex resume on a live connection

[Sandstorm831]

Parent

Spec #276; implementation sub-issue of M7 #275.

What to build

When a Turn's requested resume names the thread already live on this Codex app-server connection, and the Adapter has not marked the Session detached, the Codex Adapter skips thread/resume and sends turn/start. A reopened Run or a detached Session still resumes the exact conversation. The change sits inside the Codex Adapter behind the Harness Seam; Claude Code is unchanged (stories 112–113).

Acceptance criteria

• Codex replayer conformance proves a second human Turn on a live connection sends exactly turn/start with no thread/resume.
• Conformance proves a reopened or detached Session still resumes exactly.
• The shared conformance suite still passes against the fake.
• CI scenario: runtime conformance is green on Windows x64, macOS arm64, and Linux x64.

Blocked by

• M7: Report structural violations from one rule catalogue in a dedicated gate step #277
• M7: Report guidance-structure violations in the contract and check Module guidance shape #278
• M7: Report vendor-provenance violations in the contract #279
• M7: Report release-workflow violations in the contract #280
• M7: Enforce unused code and clear today's findings #281
• M7: Fail lint on any warning and add two OpenCode style rules #282

Implementation packet

Owner: Codex Harness Adapter, behind the Harness Seam

Ratchet: conformance for no redundant resume on a live connection

Deletes: the redundant thread/resume on a live connection

Gate: ADR 0027 three-OS CI and recorded human evidence, with the slice scenario above.

Dependencies: no new runtime dependency

[Sandstorm831]

Starting context

Files are pinned at b976b6c, which carries #281's unused-code gate. src/harness/codex.ts, src/run/execution/agent.ts, and the Codex conformance file are unchanged in behaviour since the spec's fb3ee06. Read the Authority column before using older wording.

Input	Why it matters	Authority
Spec #276 stories 112–113, Implementation Decisions › Interrupt and Steer fixes › Redundant Codex resume, and Testing Decisions › Seams 5 (Harness conformance), bullets 1–2 and its last two sentences	The target. Skip thread/resume only when the requested resume names the thread already live on this connection and the Adapter has not marked the Session detached. A reopened Run or a detached Session still resumes exactly. Claude Code is unchanged. The spec does not say how the test gets the live coordinate or which replay mode it uses.	Canonical spec
#255 decision, "No-decision fixes surfaced", item 2	The origin of the fix: a redundant thread/resume before every human Turn on a live connection. ADR 0035's redesign is not this ticket.	Canonical decision
CodexSession.submit, with the resume choice, the different-thread fence, the detached flag, and markDetached	The change goes here. Today any turn.request.resume sends thread/resume, even when it equals the live this.coordinate. markDetached is called on every lost and interrupted settle (control failure, connection end, lost acceptance, protocol corruption, interrupted terminal), so the Adapter already knows exactly which Sessions are detached. The fence for a different thread stays as it is.	Current behavior
CodexPreparedHarness.startTurn	Sessions live in a map on one prepared Harness. A reopened Run or a newly prepared Harness has no coordinate, so it must still resume from the caller's coordinate.	Current behavior
sessionRecovery, driveInteractiveTurn's detachAfterTurn, and settleTurnResult's detach override	Why the redundant resume happens. Execution records every interactive Turn's Session detached, and the next Turn carries that coordinate as resume. That stays: #298 keeps recording detached (its change there is comment only). The fix is Adapter-only. Execution keeps passing resume.	Current behavior
Measured at b976b6c with a scratch script (not a checked-in test) over the strict two-turns recording: Turn 1 completes, and Turn 2 is sent with Turn 1's session event coordinate as resume	Stdin is thread/start, turn/start, thread/resume. The strict matcher refuses the unrecorded thread/resume and exits, so Turn 2 settles failed / recovery-unacknowledged with its Session unusable. This test is red at base without a new recording, and it goes green once the Adapter sends exactly turn/start.	Current behavior
two-turns case and recording.json	Fixture read, not changed. It records thread/start, turn/start, turn/start on one thread (codex-cli 0.155.0, codex-probe-2). Re-record only if the implementer chooses a different case, and state the semantic change on the issue (testing).	Current tests
Codex conformance: later fresh Turns reuse one thread, Adapter-detached plus caller resume, newly materialized Session, detached with resume omitted, prepareDetachedCodex, and recorded Resume after Interrupt	Fixture touched: the new live-connection case goes beside the first test. It asserts exact request methods and ids, which is the prior art for "exactly turn/start". The other four prove "reopened or detached still resumes exactly" and must stay green unchanged. The recorded Resume case is same-connection: an interrupt detaches the Session, then thread/resume follows.	Current tests
Shared suite against the fake, its resume case, and the replayers' runExactThreadRecoveryCases	Acceptance says the shared suite still passes against the fake. The fake is not expected to change. The exact-thread cases also run over the Codex replayer, and they detach through a lost first Turn, so the skip never applies to them.	Current tests
Harness adapter internals › Codex, Harness › Interrupt, recovery, and cleanup, and Harness › Tests	"A detached Session requires its exact thread/resume id … no fallback" stays true. Add the live-connection skip beside it. Codex replay runs only in the standalone runtime-conformance program, never under bun test.	Current guidance
T3 Code de251fc297 apps/server/src/provider/Layers/CodexSessionRuntime.ts:L2481-L2496 (thread opened once at session start), :L2550-L2612 (each sendTurn sends only turn/start), :L743-L786 (openCodexThread)	Prior art: an app-server client opens or resumes a thread once per connection and never per Turn. Its recoverable-error fallback from thread/resume to thread/start is exactly what ADR 0022 forbids here.	Advisory

Reading rule

Start with the spec's Redundant Codex resume bullet, then CodexSession.submit and the four recovery tests.

• Skip only when turn.request.resume.opaque equals the live coordinate and the Session is not detached. Every other case keeps today's path: no coordinate yet, Adapter-detached, a different thread (fenced as recovery failure), or a Session marked unusable. Never start a fresh thread in place of a resume.
• No Harness Interface, execution, or Application change. Claude Code is not touched.
• Land the live-connection case red at base. The strict two-turns replay gives that for free, as measured above.
• M7: Refuse a failed Codex control call without losing the Turn #297 is blocked by this ticket and builds on it. It keeps a refused control call from marking the Session detached, then proves the next Turn still skips the resume. So key the skip on the Adapter's own detached flag, not on the last Turn's result kind. M7: Refuse a failed Codex control call without losing the Turn #297 then edits the same files: src/harness/codex.ts (CodexTurn control there, CodexSession here), the Codex conformance file, and harness-adapters.md (line 54 there, line 51 here).
• The unused-code gate now runs in check (baseline). Export nothing new unless another file imports it.

Open for the human

Choices the spec leaves open, each with the drafting recommendation. They are settled with the human when this context is refreshed for ready-for-agent.

1. Test design. Recommendation: use the strict two-turns recording. At base, its second Turn sends thread/resume and fails recovery-unacknowledged (measured), so the test is red with no re-record.

Owner: Codex Harness Adapter, behind the Harness Seam
Ratchet: conformance for no redundant resume on a live connection
Deletes: the redundant thread/resume on a live connection
Gate: ADR 0027 three-OS CI and recorded human evidence, with the slice scenario above.
Dependencies: no new runtime dependency