fix(qa): persist direct-API qa_fixer status via a qa_status.json artifact

[OBenner]

Problem

On a direct-API provider (e.g. OpenAI via the generic_edit runtime), the qa_fixer can't record its status, so a REJECTED → fix → re-QA loop stalls. prompts/qa_fixer.md tells the model to write qa_signoff into implementation_plan.json using Claude's Write/Edit tools and absolute paths — but in generic_edit those tools don't exist (the real tools are write_file/replace_text) and resolve_workspace_path rejects absolute paths. The result: ready_for_qa_revalidation is never set, is_fixes_applied stays false, and the fixer reports error even after applying real fixes.

Fix

Give the direct-API fixer a file-based status contract it can actually satisfy, mirroring the file-based verdict the Claude path already relies on:

1. Prompt override — build_qa_fixer_prompt gains runtime_status_relspec. When set, a generic_edit addendum overrides the Claude Write/Edit/absolute-path guidance: edit with write_file using workspace-relative paths, and record completion by writing a workspace-relative qa_status.json (status="fixes_applied", ready_for_qa_revalidation, fix_session, fixes_summary), then finish. The Claude path is unchanged (param defaults to None).
2. Deterministic merge — run_qa_fixer_via_runtime computes the spec dir relative to the project root and, when is_fixes_applied is false, folds <spec_dir>/qa_status.json into qa_signoff (load_/save_implementation_plan), then re-checks. Pre-existing signoff fields are preserved.
3. Same-loop async close — run_qa_fixer_runtime_session async-closes each per-attempt provider session in a finally (rebuilt per attempt by the recovery loop). Adds OpenAICompatibleSession.aclose() to await the underlying AsyncOpenAI connection-pool close — the sync close() only drops references, leaking sockets.

Tests

tests/test_qa_fixer_runtime.py — 18 passed (10 existing + 8 new), ruff clean:

apps/backend/.venv/bin/python -m pytest tests/test_qa_fixer_runtime.py -p no:cacheprovider -q

New coverage: scripted runtime session writes qa_status.json → returns "fixed" with the artifact merged into qa_signoff; no-artifact stays "error" (loop must not falsely pass); the prompt-addendum content; per-attempt async close; and the OpenAICompatibleSession.aclose() primitive.

Follow-up (not in this PR)

The qa_reviewer runtime path (run_qa_reviewer_via_runtime) has the same latent status-persistence bug — it also can't write qa_signoff on a direct-API provider. Its verdict is richer (approved/rejected + issues_found), so it needs the same pattern with an extended artifact schema. Tracked separately to keep this change focused.

🤖 Generated with Claude Code

[coderabbitai]

Caution

Review failed

An error occurred during the review process. Please try again later.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch claude/priceless-banach-3935b8

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[sonarqubecloud]

Quality Gate passed

Issues
3 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud