Review all

Runs every review skill installed on this machine against one change and hands back one report. Each skill looks from a different angle, so their overlap is a signal: a problem two skills find on their own is more likely real.

Input: the skill args: a PR number, a PR URL, branch with an optional base commit, or nothing for local uncommitted changes. A caller skill may add caller: plan-and-build; see step 4.

Rules for the whole run

• Report only. No skill in this run edits code, commits or posts to GitHub. Tell each skill so when you invoke it. Why: the reviewer never fixes what it reviewed (global CLAUDE.md, "Verification"). Fixes come after the user reads the report.
• One skill at a time. code-review and interrogate each fan out several agents. Running them together breaks the 5-agents-at-once cap and hits rate limits.
• A failed skill is not a clean result. If a skill errors or its agents die, re-run it once. If it fails again, list it under "Did not run" in the report. Never read silence as "no findings".

Step 1: pin the target

Work out the target once and pass the same one to every skill, so all five review the same code.

• Args hold a PR number or URL: PR mode. Record the PR number and its head commit (gh pr view <n> --json headRefOid).
• Args start with branch: branch mode, for committed work on a feature branch. If a base sha follows, BASE is that sha. Otherwise find the default branch (git symbolic-ref --short refs/remotes/origin/HEAD; exit code 128 means no remote HEAD, so fall back to whichever of main, master exists) and set BASE=$(git merge-base <default> HEAD). No default branch found: ask the user for the base. Review git diff $BASE..HEAD. No commits since the base: say "Nothing to review" and stop.
• Otherwise: local mode. Record git rev-parse HEAD and git diff HEAD --stat plus untracked files from git status --porcelain. Nothing changed: say "Nothing to review" and stop.

Every skill writes its own report files to <scratchpad>/review-all/, never into the repo. code-review normally writes .claude/reviews/; tell it to use the scratch folder instead. With caller: plan-and-build, also tell every skill: "non-interactive: ask no questions, run no hyperresearch, write only under the scratch folder". After each skill, git status --porcelain must show nothing new; revert anything a skill wrote into the repo and log it.

Step 2: run the five skills, in this order

Invoke each with the Skill tool. Pass the target from step 1 and the line "report only: do not edit code, commit or post anywhere".

After each skill, save its findings to a scratch file (<scratchpad>/review-all/<n>-<skill>.md) before you start the next one. That way the merge reads files, not a long context.

Step 3: merge into one report

1. Read the five scratch files.
2. Refute skills 2 to 5. code-review already refutes its own findings. The findings from ponytail-review, blast-radius, thermo-nuclear-code-quality-review and interrogate arrive unchecked. Batch them, about 5 per Sonnet agent, grouped by file. Each refuter tries to disprove its findings by a different method than the one that found them (run a command, or read the connected code instead of the diff). Follow code-council's refuter rules (~/.claude/skills/code-council/SKILL.md, "Accuracy first, then cost" section). Medium, high and critical findings also get a second blind refuter: a different agent that never sees the first verdict and must run something (a command or a test), not only read. Verdicts: REFUTED (drop it, and note it in the report), CONFIRMED (keeps its proof), UNVERIFIED (when unsure; never default to REFUTED). At most 5 agents at once. A dead or errored refuter is not a verdict: re-run it once, then list its findings as UNVERIFIED and the refuter under "Did not run".
3. Merge duplicates: the same defect at the same file:line, or the same root cause, is one finding. Keep the clearest wording and list every skill that found it.
4. Severity: map every skill's scale to one: critical, high, medium, low (Critical = critical, Important = high, Must fix = high, Minor or Nice to have = low). Use the highest severity any skill gave it. A finding that two or more skills reached on their own goes up one level (global CLAUDE.md rule).
5. Proof: every finding needs file:line plus a command, test or query that shows it. A finding that survived step 2 keeps that proof. Any other finding is labelled UNVERIFIED. Keep it in the report and don't drop it.
6. Conflicts: if one skill flags something and another argues it is fine, keep the finding and show both sides in one line.

Load the unslop skill, then write the report:

• Top: what was reviewed (target and commit), the verdict (ship, fix first, or rethink), and what the user must decide.
• Findings, most severe first. Each one: the problem in plain words, a concrete example, why it matters, the smallest fix (prefer fixes that delete code), and which skills found it. Evidence (file:line, command and output) goes last, under "How we know".
• Cost: one line. Count the agents you launched and the refuters yourself. Take tokens from code-review's run-cost line ("Agents: ... cost: ..."). Write "tokens unknown" for the other four skills, because they report no cost.
• Did not run: any skill that failed twice, and why.

Write the merged report to <scratchpad>/review-all/report.md. The session model does the merge; if the session is not on Opus, hand the five files to one Opus subagent to merge. Outside caller mode, if the ReportFindings tool is available, report through it instead of as text.

Step 4: ask before acting

With caller: plan-and-build, skip this step. Return to the caller: the verdict (ship, fix first, or rethink), the report path, and every critical and high finding, UNVERIFIED ones included. The caller handles fixes.

Otherwise, end by asking the user:

• which findings to fix (fixes go through a separate review afterwards), and
• in PR mode, whether to post the report to the PR. Post only on a clear yes in this conversation.