Run a rigorous, multi-dimensional review checklist on any pull request before opening, merging, or after receiving review comments. Use this skill ALWAYS before opening a PR, when asked to "review this PR/branch/diff", "check my changes", "is this ready to merge", or before promoting a branch from dev to staging or staging to main. Goes beyond style — checks security, performance, testing rigor, supply chain, and Definition of Done coverage.
Installation
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Run a rigorous, multi-dimensional review checklist on any pull request before opening, merging, or after receiving review comments. Use this skill ALWAYS before opening a PR, when asked to "review this PR/branch/diff", "check my changes", "is this ready to merge", or before promoting a branch from dev to staging or staging to main. Goes beyond style — checks security, performance, testing rigor, supply chain, and Definition of Done coverage.
Pull Request Review
This is the rubric. Apply it to every PR before opening and again before
merging. The reviewer subagent uses this checklist with read-only tools
(Read, Grep, Glob).
Model Routing
When a PR needs a cross-model second opinion, resolve the backend/security
review model through:
Default preference is Codex for backend, security, infrastructure, and edge-case
review. Frontend-heavy PRs may additionally use
walter_model_resolve frontend review_model. PHI/medical diffs must use
walter_model_resolve phi review_model and stay local.
Required before opening a PR
PR title validated: run ./hooks/pr-title-validator.sh "$TITLE" before
calling gh pr create. If exit != 0, fix the title — do not proceed.
Format: [TYPE] -CATEGORY- title (see CONTRIBUTING.md -> "Title convention").
Spec exists at docs/specs/<slug>.md and is current
Plan exists at docs/specs/<slug>.plan.md and matches the diff
All acceptance criteria in spec map to ≥1 test
Tests pass locally on the latest commit: unit, integration, e2e
No console.log, println!(), print(), dbg!() debugging artifacts
No secrets in diff (grep for typical key prefixes)
No commented-out code blocks (delete or commit-message it)
Branch flow respected per WALTER_BRANCH_FLOW (single-tier default → main; three-stage → dev). See ADR 0013.
Commit messages follow conventional commits
PR description references spec and ticket (Refs: docs/specs/...,
Closes TRI-NNN)
Code review dimensions
For every changed file, check each dimension. Score 0 (clean), 1 (minor),
2 (blocking).
1. Correctness
Edge cases: empty, null, single element, max size, unicode, negative,
zero, NaN, very large, concurrent access.
Off-by-one in loops and slicing.
Time and timezone: anywhere now(), Date, chrono, ensure UTC at
storage boundary, render in user TZ.
Number precision: f64 for money is wrong; use integer cents or BigInt.
For Solana lamports, always integer.
Async correctness: every await/? checked, no swallowed errors,
cancellation handled.
2. Security
Input validation at trust boundaries (HTTP, RPC, file, IPC).
Output encoding: HTML-escape, JSON-encode, SQL-parameterize. No string
concat into queries.
AuthN before AuthZ. AuthZ before any data return.
No secrets in logs, even at debug level. No PII either (especially for
[Project B]).
Crypto: use library primitives; never roll your own. Constant-time
comparison for sensitive checks.
For Solana code: signer checks on every account that should be signed,
owner checks on every PDA, account discriminator validation.
Dependencies: any new npm/cargo/pip dep in this PR? Check it via
daily-supply-chain-audit script: walter-os audit-pkg <name@version>.
3. Performance
Anything in a hot path (request handler, message loop, gRPC stream)
must avoid: allocations per call, unbounded loops, blocking I/O on
the async runtime, unbounded channels.
Database: every query has an explained index path. N+1 queries get
flagged. Migrations pre-checked on staging dataset size.
For Yellowstone/Geyser code: profile before merging anything in the
per-message critical path. Submit a flamegraph in the PR if changes
the path.
4. Testing rigor
Tests aren't just present — they actually fail when mutated.
Assertions are specific (toBe(42), not toBeTruthy()).
Each test asserts one thing (or one tightly grouped behavior).
Test names describe behavior, not implementation.
Slow tests (>100ms unit, >1s integration) flagged.
5. Readability
Function name says what it does, not how.
Function length: >50 lines is suspect, >100 is blocking.
Cyclomatic complexity: >10 is suspect, >15 blocking.
Comments explain why, not what. Code that needs a "what" comment
should be refactored into a well-named function.
Public APIs documented with examples.
6. Backward compatibility
API change? Versioned or deprecated, not removed.
DB migration? Reversible. Tested on staging before main.
Config schema change? Old values still read with sensible defaults
for one release cycle.
7. Operational
New env var? Documented in .env.example and docs/operations.md.
New external dependency (DB, queue, API)? Failure mode considered.
Circuit breaker or retry-with-backoff if appropriate.
New scheduled job? Idempotent. Logged with structured fields.
Observability: any new code path that fails silently is a bug. Add
metrics or logs.
Blocking findings template
When you find a blocking issue, write a comment in this format:
All blocking findings resolved (not just acknowledged — actually fixed).
All non-blocking findings either fixed or explicitly acked in PR
comment with reasoning.
The reviewer's confidence: "I would deploy this myself, on a Friday."
If not, keep iterating.
Iteration limits
Minimum: 2 review rounds (the reviewer subagent + Copilot).
Maximum: 5 rounds. After 5, escalate to operator. Either the spec is
unclear, the implementer has wrong context, or the change is too big
for one PR (split it).
What this skill does NOT do
Replace architectural review. Big design decisions require operator
input via the architect subagent before any code is written.
Replace human review on [Company] work. The work context disables
auto-PR for that reason.
Catch subtle business-logic bugs in domains the reviewer doesn't
understand. For [Project A] (procurement law) and [Project B] (medical
compliance), invoke domain-specific subagents.
