review-all

name	review-all
description	Run four independent PR reviews (/review-pr, /review, /code-review, /audit-tests) in parallel and merge them into one deduplicated, severity-tiered report. Use for a thorough multi-lens review before merging a non-trivial PR.
disable-model-invocation	true
allowed-tools	Read Grep Glob Bash(gh *) Bash(git *) Agent
argument-hint	["pr-number"]

ProbPipe Consolidated PR Review

Review pull request $ARGUMENTS (if no number is given, detect the PR for the current branch with gh pr view) by running four independent reviews in parallel and merging them into one report.

This is READ-ONLY. Do not edit any files. Present the consolidated report and wait for the user to decide what to act on.

The four lenses are complementary, and running them as separate agents keeps their judgments independent. A finding raised by two lenses is high-confidence; a disagreement between lenses is itself signal.

Step 1: Gather context

gh pr view <number> --repo TARPS-group/prob-pipe --json title,baseRefName,headRefName,files,state
gh pr diff <number> --repo TARPS-group/prob-pipe

Note the head/base branches, the changed files, and a one-line statement of what the PR is meant to do. Pass this to all four agents so none re-derives it.

Step 2: Spawn the four reviews as parallel agents

Spawn all four in a single message (four Agent tool calls, general-purpose type) so they run concurrently. Each agent invokes the corresponding skill on the PR and returns its report. In every prompt, include the Step 1 context and require: READ-ONLY; do not spawn further sub-agents; return a self-contained structured report with no preamble or sign-off (it goes to a coordinator for consolidation).

• Agent 1 — invoke the review-pr skill (ProbPipe conventions, docs, tests, duplicate code, AI-artifact comments).
• Agent 2 — invoke the review skill (general correctness, conventions, performance, test coverage, security).
• Agent 3 — invoke the code-review skill (focused, high-effort correctness bug-hunt of the diff; it may run code read-only to confirm/refute findings).
• Agent 4 — invoke the audit-tests skill, scoped to the test files the PR touches (test gaps, weak or tautological assertions, stale/duplicate tests, and the mathematical correctness of the assertions). If the PR changes no tests, it judges whether the changed behavior should have had coverage.

If a skill is unavailable in the agent's context, tell it to fall back to performing that review directly per the lens description above.

Step 3: Consolidate

Merge the four reports into ONE — do not concatenate.

1. Deduplicate findings that describe the same underlying issue, even when worded or located differently. Note where lenses agree independently.
2. Re-tier by consensus severity (your judgment, informed by how many lenses flagged it and how load-bearing it is): Must fix / Should fix / Minor / polish.
3. Attribute each finding to the lens(es) that raised it; keep file:line.
4. Surface disagreements explicitly, with a reconciliation grounded in STYLE_GUIDE.md / CONTRIBUTING.md.
5. Keep a "Verified correct" section carrying forward what the bug-hunt and test-audit lenses checked and found sound.
6. End with a recommendation — what you would do, in order — and offer to implement it.

Output format

# Consolidated Review: PR #<n> — <title>

## Verdict
## Must fix
## Should fix
## Minor / polish
## Contested (lenses disagreed)
## Verified correct
## Recommendation

Notes

• Four agents is real token cost. For a quick look, /review-pr alone is enough; reach for /review-all before merging a non-trivial PR.
• Explicit-invocation only — general "review this PR" requests route to /review-pr.