review-core

Files (8)

• ADVERSARIAL.md
• ALTERNATIVES.md
• BLIND-REVIEW.md
• COMMIT-HYGIENE.md
• EXPLORATION.md
• README.md
• SKILL.md
• TRIBUNAL.md

SKILL.md (first lines)

# Review core

The shared review engine: **blind** critics → **tribunal** → **adversarial** validation. Predictability over improvisation.

Parent skills usually invoke review-core; direct local branch review is also supported.

**Requires sibling install** of this skill when using [review-tribunal](../review-tribunal/SKILL.md) or [ticket-to-pr-pipeline](../ticket-to-pr-pipeline/SKILL.md).

## Mode

Parents must set **mode** before running:

| Mode | Parent | Outcome for Holds |
|------|--------|-------------------|
| `review` | review-tribunal Steps 3–5 | Adjudicated findings → parent delivery |
| `implement` | ticket-to-pr-pipeline Phases 3–4 | Code fixes + re-run tests on branch |

## Code exploration

All phases and subagents follow [EXPLORATION.md](EXPLORATION.md). Paste those rules into every critic and validator prompt.

Before Step 1, the caller fills this source manifest and reuses it unchanged across critics, validators, and scorecards:

```text
Target: {TARGET}
Repo: {FULL_REPO_PATH}
Branch: {BRANCH_OR_DETACHED}
Head: {HEAD_SHA}
Diff source: {EXACT_PR_RANGE_COMMIT_OR_WORKTREE_SOURCE}
Working-tree components: {STAGED | UNSTAGED | UNTRACKED | NONE}
Commands: {EXACT_READ_ONLY_SOURCE_COMMANDS}
Changed files: {DEDUPLICATED_LIST}
Commits in range: {COUNT | 0}
Commit SHAs: {LIST | N/A}
Commit commands: {commands that produced the list | N/A}
```

## Step 1 — Blind critics

Spawn parallel **blind** judges — ticket + diff + intel facts only; no implementer rationale.

Details: [BLIND-REVIEW.md](BLIND-REVIEW.md).

- Pick distinct lenses from the roster; floor ≥2 independent critics.
- **Alternatives critic is required** for `mode: review`; default for `mode: implement` unless diff is trivial (note skip on scorecard).
- Manifest lists commits → spawn hygiene critic per [COMMIT-HYGIENE.md](COMMIT-HYGIENE.md).
- Paste [EXPLORATION.md](EXPLORATION.md) rules into every subagent prompt.

**Completion criterion:** ≥2 independent reviews returned; wrong-diff reviews discarded; skipped roster lenses noted with reason.

## Step 2 — First tribunal

You are the judge, not the author's advocate. For every material finding: **Holds up** / **Reject** / **Defer**.

Details: [TRIBUNAL.md](TRIBUNAL.md). Alternative proposals: [ALTERNATIVES.md](ALTERNATIVES.md).

Verify claims with **Read** / **Grep** / **Glob** before ruling.

Publish a scorecard to the user (finding → verdict → why). In `review` mode, hand findings to the parent only after adversarial validation; in `implement` mode, wait for validation before changing code.

**Completion criterion:** every material finding has an explicit verdict; user can see the audit trail.

## Step 3 — Adversarial validation

Spawn a **second** wave tasked with **disproving or downgrading** the **Holds up** items; confirmation must overcome counter-evidence.

Details: [ADVERSARIAL.md](ADVERSARIAL.md).

Feed validators: the finding, first-tribunal rationale, ticket/AC, sibling-repo paths, and any Atlassian/FE evidence already gathered.

Re-adjudicate after validators return. Overturn freely when evidence wins.

**Completion criterion:** each prior Hold has a CONFIRMED / DOWNGRADE / REJECT / UPGRADE verdict with evidence; revised scorecard published.

## After review-core

| Mode | Next step |
|------|-----------|
| `review` | review-tribunal target-specific delivery |
| `implement` | Implement Holds, re-run tests, summarize audit trail |

## Checklist

```
- [ ] Blind critics (≥2 distinct lenses; alternatives unless trivial skip noted; hygiene run or skip noted)
- [ ] First tribunal scorecard
- [ ] Adversarial validation + revised scorecard
- [ ] Parent handoff (target-specific review delivery or implement fixes)
```

From lucavb/skills, MIT.

More from lucavb/skills