task-observer: deep-review updates from 2026-05-06#10
Conversation
Add the inverse of the violation-escalation rule: when AGENTS.md explicitly de-scopes a capability (e.g. "No ARIA work yet", "desktop-only", "no-i18n"), findings targeting that area are inadmissible and must not be auto-fixed. Also note de-scoped concern check in the UI reviewer accessibility criteria so subagents apply the same suppression.
![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: 7909c4c867
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".
| - **If the independent reviewer is unavailable or fails**, proceed with the orchestrator-side reviews. Note which independent reviewer was skipped in the summary. | ||
| - **If no frontend files changed**, skip the UI review entirely. Don't mention it in the summary. | ||
| - **Respect project conventions.** Check CLAUDE.md for project-specific rules and flag violations. | ||
| - **Respect project conventions.** Check CLAUDE.md/AGENTS.md for project-specific rules and flag violations. Apply the inverse equally: if AGENTS.md explicitly de-scopes a capability ("No ARIA/screen reader work yet", "desktop-only", "no-i18n", "single-user"), findings targeting that capability are inadmissible — skip them and do not auto-fix. One-directional calibration (escalating violations without suppressing de-scoped concerns) produces out-of-scope changes the user didn't ask for. |
There was a problem hiding this comment.
This new rule makes de-scoped findings inadmissible only when AGENTS.md declares the de-scope, but Step 1 in the same workflow instructs reviewers to derive rules from CLAUDE.md or AGENTS.md. In projects that declare scope limits in CLAUDE.md (or docs it references), reviewers following this text can still raise and auto-fix out-of-scope items, which defeats the purpose of this change. Include CLAUDE.md-sourced de-scopes in the inadmissibility rule so scope suppression is consistent.
Useful? React with 👍 / 👎.
2 participants
Target
skill:
.agents/skills/deep-review/SKILL.mdObservations addressed
Generated by Claude Code