.github: add low-risk label trigger for Claude review with approval

[hieblmi]

Summary

Its up to the PR creator to apply the low-risk label after judging the PR impact.
The first PR reviewer(besides claude) must then assess the severity category.
If they aggree claude's green check counts.

• Adds a claude-approve job triggered by the low-risk label
• Claude reviews the PR and approves with gh pr review --approve only if no significant issues are found
• If issues are found, Claude leaves a comment instead of approving
• The approve job has pull-requests: write permission (scoped only to this job)
• Existing claude-review job (comment-only) remains unchanged

Triggers

Trigger	Job	Can approve?
@claude mention	claude-review	No (read-only)
claude-review label	claude-review	No (read-only)
low-risk label	claude-approve	Yes (if no issues found)

Test plan

• Add low-risk label to a clean PR — Claude should review and approve
• Add low-risk label to a PR with issues — Claude should comment without approving
• Verify @claude mention and claude-review label still work as comment-only

[gemini-code-assist]

Note

Gemini is unable to generate a summary for this pull request due to the file types involved not being currently supported.

[starius]

LGTM!

Some proposals added.

[starius]

s/CLAUDE.md/AGENTS.md/

[starius]

Does it have access to secrets.CLAUDE_CODE_OAUTH_TOKEN?

Codex thinks that we need to add this so it had access to it in PRs from forks:

   pull_request:
     types: [labeled, synchronize]
+  pull_request_target:
+    types: [labeled, synchronize]

[hieblmi]

Yes we have access, the @claude wouldn't work otherwise

[starius]

Then we need to replace pull_request with pull_request_target here

[starius]

s/CLAUDE.md/AGENTS.md/

[starius]

The bot can request changes to undo its previous approval:

            If NOT approving:
              gh pr review ${{ github.event.pull_request.number }} --request-changes --body "Claude review: not eligible for low-risk auto-approval. Include intrinsic PR risk and findings w
ith severities."

[hieblmi]

Good point!

[starius]

I propose to separate intrinsic severity of the PR (e.g. whether it touches any code working with funds) and severity of potential problems found in the PR (where or not PR itself is valid). If any of these is higher than low, do not use auto-approve flow.

Here is the whole prompt:

REPO: ${{ github.repository }}
PR NUMBER: ${{ github.event.pull_request.number }}

Review this pull request thoroughly, checking for:
- Code quality and best practices
- Potential bugs or issues
- Performance considerations
- Security concerns
- Test coverage

Use the repository's AGENTS.md for guidance on style and conventions.

Classify the PR's intrinsic risk severity using one of:
- critical: must block merge
- high: large blast radius or very sensitive code paths
- medium: meaningful production or maintainability risk
- low: constrained blast radius and straightforward rollback

Intrinsic risk is about the impact and sensitivity of changed code,
even when no concrete bug is found.
Any changes touching fund movement, signing/sweeping, swap state
transitions, security/authz/authn logic, secrets handling, or DB
schema/migrations are at least medium risk.

Also classify each finding using one of:
- critical
- high
- medium
- low
- nit

Decision rule:
- Approve ONLY if intrinsic PR risk is low AND highest finding
  severity is low or nit.
- If intrinsic PR risk is medium/high/critical, DO NOT approve and
  submit a changes-requested review.
- If any finding is medium/high/critical, DO NOT approve and submit
  a changes-requested review.
- If uncertain, treat as medium risk.

If approving:
  gh pr review ${{ github.event.pull_request.number }} --approve --body "Claude review: intrinsic PR risk low and findings low/nit; safe for low-risk path."

If NOT approving:
  gh pr review ${{ github.event.pull_request.number }} --request-changes --body "Claude review: not eligible for low-risk auto-approval. Include intrinsic PR risk and findings with severities."

[hieblmi]

Thank you for the great refinement.