[zbigniewsobiecki]

Problem

When a completion check fails after the first turn (e.g. a respond-to-pr-comment run that answered a question without pushing a commit), the Codex backend resumes the recorded thread. The generated command was:

codex exec resume <id> --json … -C <dir> -m <model> -s <mode> …

-C/--cd and -s/--sandbox are parent-only codex exec options (not clap globals), so the pinned CLI rejects the whole invocation before contacting the model:

error: unexpected argument '-C' found
Usage: codex exec resume [OPTIONS] [SESSION_ID] [PROMPT]

The run is then recorded as failed even though the first turn's work (e.g. the posted PR comment) already landed. Introduced in #1521, whose unit test pinned the broken ordering against a mocked spawn.

Fix

buildArgs now emits every option first and the subcommand last:

codex exec --json … -C <dir> -m <model> -s <mode> … resume <id> -

Verification

• Fed the real buildArgs output to @openai/codex@0.145.0 (the worker pin): it now fails only at thread lookup (no rollout found for thread id …, exit 1), not in argument parsing. Same grammar holds on 0.155.1.
• Unit tests pin the structure (exec first, option set identical to the initial turn, resume <id> - last) instead of the old ordering.
• npm run verify green.

Follow-ups (separate PRs)

• Codex bump to 0.155.1 + gpt-6-astra in the model dropdown (with a build-time parser smoke check).
• Spec 025: outcome-aware completion so a comment-only PR reply completes without a push.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[nhopeatall]

Summary

APPROVE — correct, minimal bug fix. buildArgs now emits all codex exec options before the resume subcommand, matching the pinned CLI grammar (codex exec [OPTIONS] resume [SESSION_ID] [PROMPT]) so continuation turns are no longer rejected with error: unexpected argument '-C' found.

Verification

• Fix is correct: resume <sessionId> moves from position 2 (immediately after exec) to just before the trailing - stdin marker. Parent-only exec options (-C, -s, -m, -o, --output-schema, -c …) now precede the subcommand, and the - PROMPT positional stays last — the prompt is delivered via child.stdin.write(promptText) regardless of turn, so the stdin mechanism is unchanged.
• No regression on the initial turn: with sessionId undefined the output is exec --json … -, byte-for-byte identical to the prior behavior (the old resume insertion was already conditional).
• Resume path is real: runState.sessionId is populated from the thread.started event (src/backends/codex/index.ts:397) and fed to buildArgs on continuation turns.
• Tests: ran tests/unit/backends/codex.test.ts → 104/104 passing. The new keeps every exec option ahead of the resume subcommand test asserts the resume option set equals the initial-turn option set and that resume <id> - is last; the CodexEngine mock also confirms -o output still resolves in the resume invocation.
• Docs: CHANGELOG.md and docs/architecture/05-engine-backends.md accurately describe the new ordering.

No blocking or should-fix issues. CI was still in progress at review time; the code change itself is sound and locally verified.

🕵️ claude-code · claude-opus-4-8 · run details

[nhopeatall]

Summary

APPROVE — correct, minimal fix that reorders the Codex resume invocation so every parent-only codex exec option (-C/--cd, -s/--sandbox, and the rest) precedes the resume subcommand, with tests that pin the right structure and strong verification against the pinned CLI.

Verification

• Fix is sound and isolated. buildArgs is the only constructor of the resume command (rg 'resume' across src/ confirms), so there is no second code path still carrying the old exec resume <id> … ordering. The - stdin prompt marker stays last and the child is spawned with stdio: ['pipe','pipe','pipe'], so prompt delivery is unchanged.
• No regression risk. The prior ordering was rejected at argument parsing (error: unexpected argument '-C'), i.e. resume never worked at all; the new ordering is a strict improvement and the author validated it parses on both @openai/codex@0.145.0 (worker pin) and 0.155.1, failing only at thread lookup.
• Tests are meaningful. resumeArgs.slice(1, resumeIndex) === initialArgs.slice(1, -1) asserts the resume turn carries the identical exec-level option set as the initial turn — exactly the invariant that #1521 broke. Ran vitest run --project unit-backends tests/unit/backends/codex.test.ts → 104/104 pass. CHANGELOG + 05-engine-backends.md note accurately describe the new grammar.

Non-blocking note

• The PR bundles a second, unrelated commit (b98c10d6) bumping js-yaml 4.3.1 → 4.3.2 in dependencies + overrides to clear a high-severity npm audit (GHSA-2883-xcg3-v3hh). The change itself is clean and consistent (js-yaml is a real direct dep used across src/), but it isn't mentioned in the PR description — worth a one-line callout there for reviewers/changelog traceability. Not a blocker.

🕵️ claude-code · claude-opus-4-8 · run details