[shahar1]

Summary

• A review thread GitHub marks outdated blocked APPROVE until someone resolved it, even when a later push had already fixed what it asked for. On a base branch with required_conversation_resolution, the PR then sat blocked and the skill never said why. This came up for real on Add EKS operator for commands in existing Pods airflow#72542: an outdated kubernetes_conn_id thread, fixed and answered "Done", was one of the two threads that held the merge.
• code-review context now lists unresolved outdated threads (outdated_threads). The agent checks each one against the current diff (new classifications/outdated-thread.md). For each thread the code addresses, disposition --addressed-thread <id> removes it from the thread gate and returns it in resolve_first, and the skill tells the maintainer: "resolve it to unblock the merge".
• Guard rails: only threads that are both unresolved and outdated can be passed; any other id exits with an error. The author saying "Done" does not count; the current diff has to show the fix. The skill never resolves a thread itself, since that stays the maintainer's sign-off. Addressed threads never appear in the review body. Threads that are not outdated still gate APPROVE as before.

Type of change

• Skill change (.claude/skills/<name>/) — eval fixtures updated below
• Python package (tools/*/ with pyproject.toml)
• Tool / bridge contract (tools/<system>/*.md)
• Groovy reference impl
• Cross-cutting (RFC, AGENTS.md, sandbox, privacy-LLM)
• Documentation (docs/, README.md, CONTRIBUTING.md)
• Project template (projects/_template/)
• CI / dev loop (prek, workflows, validators)
• Other: cr-pr.graphql reads thread id / isOutdated and comment url (vetted-ops query, plugin copy synced)

Test plan

• prek run --all-files passes
• For Python packages touched: uv run pytest / ruff check / mypy passes — tools/pr-management/tests/code_review: 141 passed, including 3 new tests: test_step6_addressed_outdated_threads_go_to_the_maintainer, test_outdated_thread_the_code_addresses_unblocks_approve, and test_addressed_thread_must_be_outdated_and_unresolved
• For Groovy bridges touched: command-line invocation tested end-to-end
• For skill changes: eval suite passes for the affected skill — new step-6-outdated-thread suite, 3/3 with claude -p
• For skill behaviour changes: a new or updated eval fixture is included in this PR (step-6-outdated-thread: fixed as asked / code moved but the concern remains / author says "Done" but the diff disagrees)
• Other: ran the edited cr-pr.graphql live against Add EKS operator for commands in existing Pods airflow#72542. With the thread state from before it was resolved, context listed the outdated kubernetes_conn_id thread, and --addressed-thread dropped the thread gate from 2 to 1 and returned it in resolve_first. The remaining thread is not outdated, so it still blocked APPROVE.

RFC-AI-0004 compliance

• HITL — no new mutation; resolving a thread is left to the maintainer, the skill only tells them
• Vendor neutrality — placeholders kept; eval fixtures are synthetic (acme)
• Write-access discipline — no autonomous outbound messages; resolve_first is maintainer-facing only, never in the review body

Linked issues

Refs #1585 (the single APPROVE definition this narrows).

Notes for reviewers (optional)

• Scope is outdated threads only. A thread that is not outdated but already answered (for example, a reviewer withdrew their concern in a reply) still gates APPROVE. Extending the same flow to those is a separate decision, because nothing in the diff can show it is addressed.
• resolve_first is returned even when another gate keeps the result at COMMENT, so the maintainer still learns which threads they can resolve.
• SKILL.md Golden rule 7 gained one sentence; measured_tokens was restamped (3770 → 3826); surface_hash is unchanged.

---

Generative AI: drafted with Claude Code (Claude Opus 5.5), reviewed by the author.

🤖 Generated with Claude Code