[MarkusNeusinger]

Summary

The owner's call after #12470: code quality comes before saving tokens, so Gemini reviews its own finished work before Claude sees it.

• New section 2, step 7 in the gemini-task skill. Once a task (a single brief or one in packages) or a review round is done and the gates pass, Gemini pushes and starts one fresh subagent as the reviewer, in its own worktree on a detached checkout of the task branch. The reviewer's prompt names only the brief's Constraints and Acceptance and the diff, not Gemini's plan or notes, so it judges the code rather than the intent. It reads the whole diff for defects, runs the mutation table plus two or three breaks of its own, and reports findings with path:line, severity, and a fix. It commits and pushes nothing. Gemini fixes what holds up, reruns the gates, and pushes. A hand-over on a blocker skips the step.
• ## Self-review in the hand-over PR body lists every finding as fixed (with its commit) or declined (with the reason), or as not run after a blocker hand-over. Review-round comments carry the findings too.
• Claude's review stays independent (section 3, step 3). A reviewer of the same model shares the author's blind spots: in test(app): add render tests for palette, debug, and map sections #12470 the copied page logic passed Gemini's own gates. So the Opus subagent never sees the self-review, and afterwards Claude judges each declined finding. This matches the existing /zweitmeinung rule that a Gemini review never counts as the independent review of Gemini's work.
• The hand-over moves from step 7 to step 8, and all references follow.

Test plan

• uv run python -m tools.changelog check --base origin/main: the fragment is well-formed and the diff passes.
• grep for section 2, step and step 7 across the skills and docs: every reference names the new numbers.
• No automated check covers skill prose; the next Gemini task is the real test.

🤖 Generated with Claude Code

[Copilot]

🟡 Changes recommended

Three unresolved findings in .claude/skills/gemini-task/SKILL.md remain open, and the main gap is that the self-review step skips single-brief tasks even though the PR description and changelog say it runs for any finished task.

3 open issues

• Run self-review for single-brief tasks · New
• Give reviewer the brief to assess scope · New
• Add a fixed self-review entry for blocker hand-overs · New

What changed in this PR

This PR changes the gemini-task skill so Gemini has its finished work reviewed by a fresh subagent before it hands the work over. Claude's independent Opus review stays in place, because a reviewer of the same model shares the author's blind spots. The skill is Claude's process document for briefs that Gemini works in the owner's interactive agy session.

Changes:

• Adds step 7 to the Gemini section of the skill. Gemini starts a reviewer subagent that sees only the Acceptance section and the diff, and reports findings without committing or pushing.
• Adds a ## Self-review section to the hand-over PR body. Each finding is listed as fixed (with its commit) or declined (with the reason).
• Renumbers the hand-over from step 7 to step 8 and updates all references to it. Section 3 now says Claude judges each declined finding.

File	Description
changelog.d/​gemini-self-review.md	Adds the changelog fragment for the self-review step.
.claude/​skills/​gemini-task/​SKILL.md	Adds the reviewer step, the Self-review section in the PR template, and the renumbered hand-over references.

_{Review effort: Lite}