feat: multi-agent PR review panel (task #0047)#63
Conversation
Split the monolithic code-reviewer into 5 focused specialist agents. Complexity heuristic in evolve.md Step 8 determines review depth: docs-only/small PRs get 1 reviewer, medium 3, complex 4-5 in parallel. All must pass to merge. (task #47)
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9cd7a2173d
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| # DOCS-ONLY (.md only, no .py/.sh/.json/.toml): | ||
| # -> 1 agent: code-reviewer (fast path -- reports PASS immediately) |
There was a problem hiding this comment.
The new docs-only tier sends any .md-only PR to code-reviewer, and that reviewer now immediately returns PASS for docs-only changes. A prompt-only change (for example in docs/prompt/) satisfies this docs-only condition, so it can bypass the specialized reviewers that are supposed to guard autonomous-loop and documentation integrity. This creates a real path where behavior-defining prompt edits get effectively no substantive review; make docs-only exclude meta surfaces (like docs/prompt/) or evaluate higher-risk triggers before this fast path.
Useful? React with 👍 / 👎.
| # MEDIUM (100-300 lines, OR new .py module, OR new dependency): | ||
| # -> 3 agents in parallel: code-reviewer + safety-reviewer + docs-reviewer | ||
| # | ||
| # COMPLEX (>300 lines, OR 5+ files, OR touches scripts/ or docs/prompt/): |
There was a problem hiding this comment.
The complexity gate escalates to the full reviewer panel for scripts/ and docs/prompt/, but not for .claude/agents/. As written, a small PR that edits only agent definitions can be classified as SMALL and reviewed only by code-reviewer, even though the meta reviewer is the one tasked with checking agent-scope overlap and autonomous-pipeline integrity for those files. Add .claude/agents/ to the escalation trigger (or make meta-reviewer unconditional for that path) so policy-agent changes always get the intended review.
Useful? React with 👍 / 👎.
- scripts/watchdog.sh: auto-restarts daemon on crash, rate-limits
to 5 restarts/hour, sleeps 1 hour if fundamentally broken
- Restored all 37 wontfixed tasks — the agent does the work, not skip it
- evolve.md Step 6o: removed queue cap, replaced with judgment-based rule
("create only what matters, never to fill a quota")
- Merged #74 into #63 (genuine duplicate)
1 participant
Summary
Test plan
make checkpasses (663 tests, mypy clean, ruff clean)gh pr diff --statto determine tier