enterprise-pr-review

Files (3)

• SKILL.md
• evals.json · evals/
• pre-merge-review-sweep.cjs · scripts/

SKILL.md (first lines)

# Enterprise PR Review

## Repo Overlay Bootstrap

When the current git root does not yet have repo-local enterprise overlays, run
the public installer from this repository before PR review work:

```bash
node scripts/install-agentic-platform.cjs --target <repo-root> --codex-home <codex-home>
```

This creates or refreshes `.codex/repo-skills/<repo-slug>-enterprise-pr-review`,
`.codex/repo-skills/<repo-slug>-enterprise-stack-review`, the repo profile, trap
file, and overlay manifest for whichever repo is active. After it runs, prefer
the repo-local `<repo-slug>-enterprise-pr-review` overlay for repo-specific rules.

## Global Precheck

Before reading further, writing artifacts, delegating, or changing files, run:

```bash
enterprise-precheck --skill enterprise-pr-review
```

If it exits non-zero, stop and report stderr verbatim. Do not hand-craft packet files or evidence markers to bypass it.

## Overview

This skill turns PR review feedback into a closed, verified engineering loop. It is not a comment-cleanup skill. Treat every unresolved blocking thread as a possible production bug until current code, tests, CI, and review evidence prove otherwise. Treat non-blocking Copilot/advisory comments as valuable review signal to harvest, classify, and learn from without making them default merge blockers.

Use this when the user says:

- "PR 123 has unresolved conversations"
- "address Copilot comments"
- "cannot merge because conversations are unresolved"
- "learn from this failed PR review"
- "skip Copilot conversations but learn from them"
- "harvest non-blocking Copilot comments"
- "resolve PR feedback"
- "make a harness skill that handles PR conversations"

## Modes

Use exactly one mode per run:

- `blocking-closeout`: for merge-blocking unresolved conversations, requested PR cleanup, or PRs that cannot merge. Blocking threads must be fixed/replied/resolved or explicitly reported as blocked before declaring merge-ready.
- `advisory-harvest`: for Copilot/GitHub comments that are valuable but not blocking after the repo's async-review wait and thread sweep have passed. Advisory comments do not block merge by default, but every useful finding must become one of: immediate fix, tracked follow-up, trap-matrix update, repo gate/script recommendation, contract/plan rule, or skill eval.

Escalate any advisory comment to `blocking-closeout` or an immediate hotfix if it credibly indicates security, data loss, tenant isolation, schema/query correctness, money/order/invoice/inventory corruption, or broken user-visible behavior in the current diff.

## Draft PR Policy

Do not make every PR draft by default. Use this policy:

- `normal`: open PR and keep it open until required gates, async review checks, review-thread state, and any repo-specific pre-merge sweep pass. Do not enable auto-merge by default.
- `high-risk`: draft PR until enterprise review/forge/verify/harness have passed and Copilot has either reviewed the current head or an explicit advisory-harvest follow-up is recorded.
- `draft-default`: only use draft for all PRs if the repo is confirmed to run Copilot review on draft PRs. If that setting is unknown or disabled, draft-by-default can hide the PR from useful review and slow delivery.

High-risk means schema/query/data-sensitive work, tenant/security, money, orders, invoices, inventory, external integrations, destructive writes, authentication, or UI/PDF/file workflows that affect real operators.

If a PR is already open and later becomes high-risk, do not churn the PR state mechanically. Instead, record the risk, run the required gates, request/re-request Copilot review where useful, and route advisory findings through `advisory-harvest`.

## Universal Enterprise Hardening Baseline

- Use the GPT-5.5 fresh baseline: outcome, success criteria, constraints, evidence, and stop rules come before legacy prompt ceremony.
- Green checks are not enough for blocking-closeout. Async review comments can arrive after checks pass, so live blocking thread state is the merge blocker source of truth.
- Non-blocking Copilot/advisory comments are not default merge blockers, but they are not disposable. Harvest them into fixes, follow-ups, traps, gates, contract/plan rules, or evals.
- Never resolve a review thread just because it is outdated, annoying, or probably fine. Reply with evidence first, then resolve only after proof.
- If a thread identifies a real bug class, add or update a regression test before or with the fix. Do not rely on mock-only, migration-only, or diff-only proof for schema/query/data-sensitive changes.
- If the fix touches UI, PDF upload, file upload, preview/download, modal, navigation, rendered output, print labels, or browser behavior, run headless browser proof.
- If the fix touches database shape, tenant isolation, query semantics, money, order, invoice, inventory, integrations, or writes, use current runtime code reads plus live or migrated-integration DB proof where available.
- All verification must be headless. Manual browser or GitHub UI observation is supporting context only.
- If any code, contract, docs, or gate changes are made after review/forge/verify evidence, prior evidence for affected files expires.
- Keep PR review orchestration inside the current harness. Do not route enterprise PR review through external bridge roles.

## Entry Gate

Before doing PR work, the global precheck above must already have passed.

Then collect the current PR facts:

```bash
gh pr view <PR> --json url,state,isDraft,baseRefName,headRefName,headRefOid,mergeStateStatus,reviewDecision
gh pr checks <PR> --watch=false
```

Use GraphQL or the GitHub review-thread API to list review threads with:

- thread id
- resolved state
- outdated state
- file path and line
- top-level comment database id
- author
- body
- replies

Do not trust the PR page summary or stale transcript memory for unresolved count.

## Deep Think Precheck

Before fixing anything, state a compact precheck:

- Ask: PR number and requested outcome.
- Mode: `blocking-closeout` or `advisory-harvest`, and why.
- Source read: PR metadata, live thread list, current head SHA, changed files, checks.
- Root objective: close review blockers without hiding real bugs.
- Blast radius: changed runtime/doc/test/gate surfaces and required proof.
- Ownership: source-of-truth files/modules/contracts affected by each thread.
- Edge cases: stale/outdated thread, async new review, branch behind base, CI rerun, false positive, docs path deletion, test-only branch, UI/DB surface.
- SRP:
  - `fix-now`: real review findings and missing regression proof.
  - `follow-up`: broader gate improvements not required to unblock the PR.
  - `note-only`: historical comments or unrelated docs.
- Verification: local commands, repo gates, live thread re-query, CI status, head SHA, and merge readiness.

## Workflow

### 1. Snapshot The PR

From anotherben/agentic-governance, MIT.

More from anotherben/agentic-governance