<EXTREMELY-IMPORTANT> This skill is review-first and only becomes mutating after explicit user approval.

Non-negotiable rules:

1. Read the full diff before reporting any issue.
2. Load and respect checklist.md, including suppressions.
3. Stay scoped to the diff against the review base.
4. Stay read-only unless the user explicitly picks "Fix it now" for a blocking issue.
5. Keep findings terse, evidenced, and specific.

</EXTREMELY-IMPORTANT>

Branch Review Before PR

Inputs

• $request: Optional focus area, branch risk hint, or review scope clarification

Goal

Review the current branch against the default base and:

• find structural defects that survive context checks
• separate blocking issues from non-blocking issues
• ask the user how to handle each blocking issue
• optionally apply only the fixes the user explicitly approves

Step 0: Resolve the review base

Determine:

• the current branch
• the default review base, usually origin/main or origin/master
• whether the user wants full-branch review or a narrower scope

If the current branch is already the base branch, or there is no diff, stop and say so clearly.

Success criteria: There is an explicit diff base and actual branch delta to review.

Step 1: Load the checklist and diff surface

Read:

• checklist.md
• the full diff against the base branch
• changed-file list when the diff is large

Rules:

• do not review from --stat alone
• do not report checklist items that are suppressed
• when the diff is large, read changed files individually with enough surrounding context to verify claims

Success criteria: The checklist is loaded and every changed file that matters has been read.

Step 2: Run the two-pass review

Pass 1 is blocking-only:

• query and data safety
• race conditions and concurrency
• auth and trust boundaries

Pass 2 is non-blocking:

• all remaining checklist categories

For each candidate issue, verify:

• it is actually in scope
• it is not already fixed elsewhere in the diff
• surrounding code does not already mitigate it

Success criteria: All meaningful checklist categories were evaluated and false positives were filtered out.

Step 3: Report all findings

Always report the full result set:

• blocking issues first
• then non-blocking issues
• explicit clean result if none were found

Use this structure:

Branch Review: N issues (X critical, Y informational)

CRITICAL (blocking):
- [file:line] Problem description
  Fix: suggested fix

Issues (non-blocking):
- [file:line] Problem description
  Fix: suggested fix

Success criteria: The user can tell immediately whether the branch is blocked and why.

Step 4: Handle blocking issues one by one

For each blocking issue, use a separate AskUserQuestion with:

• the issue and location
• the recommended fix
• these options:

- Fix it now - Acknowledge - False positive

If the user chooses:

• Fix it now: apply only the approved fix
• Acknowledge: leave code unchanged and record the acceptance
• False positive: leave code unchanged and note the dismissal

Do not batch multiple blocking issues into one question.

Success criteria: Each blocking issue has an explicit user decision.

Step 5: Apply approved fixes only

When the user chooses Fix it now:

• make the smallest correct code change
• keep changes scoped to the approved issue
• do not commit, push, or create a PR
• summarize what changed after the fixes are applied

If the user only acknowledges or dismisses issues, make no edits.

Success criteria: Only explicitly approved fixes were applied.

When To Load References

• checklist.md

Use for the detailed review matrix, suppressions, and issue categories.

Guardrails

• Do not add disable-model-invocation; this skill must remain usable inside PR/ship workflows.
• Do not review unchanged files beyond what is needed to verify context.
• Do not report style preferences as issues.
• Do not fix anything unless the user explicitly approved that issue.
• Do not commit, push, or open a PR from this skill.

Output Contract

Report:

1. the review base and branch scope
2. blocking findings
3. non-blocking findings
4. per-blocker user decisions if questions were asked
5. any fixes actually applied