Review

Skill MakFly/ai-dev-kit/.claude/skills/review

Spec-Driven Agentic OS for Claude Code + OpenAI Codex CLI. Native skills, subagents, hooks, and an adversarial codex review loop. Inspired by github/spec-kit, built on Claude Code primitives. No daemon, no vendor lock-in.

Install

npx -y skills add MakFly/ai-dev-kit --skill review

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use PROACTIVELY after /plan, /tasks, or /implement completes — and any time the user invokes /review <slug> [artefact]. Runs both the `verifier` subagent (forked context, Claude-native) and codex (via `codex exec`) on the named artefact. Writes specs/<slug>/reviews/{verifier,codex}-<artefact>-<date>.md. Never edits the implementation.

SKILL.md

3.3 KB, as published. Nobody here has run it

/review

Dual review: native verifier subagent + adversarial codex exec challenge.

Procedure (runs inside the forked `verifier` subagent)

Resolve artefact. First arg = slug. Second arg = artefact name (spec, plan, tasks, all). Default = the most recently-modified of the three.
Native verifier work. As the verifier agent in fork mode, perform your normal job — re-run claimed checks, cross-check spec/plan/tasks/code. Write specs/<slug>/reviews/verifier-<artefact>-<ISO>.md.

Codex challenge (Bash). Run codex via stdin (avoids YAML/option parsing issues) and redirect output to file (codex --sandbox read-only cannot write to disk itself):

ISO=$(date -u +%Y%m%dT%H%M%SZ)
PROMPT_TEXT=$(cat <<EOF
You are Codex acting as an adversarial reviewer for ai-dev-kit v5.
Target artefact: specs/<slug>/<artefact>.md
Constitution: memory/constitution.md
Latest digest: supports/digests/digest-*.md (newest)

Surface, with file:line citations:
1. Contradictions between artefact and constitution.
2. Missing or non-testable acceptance criteria.
3. Optimistic claims without evidence.
4. Risks omitted from plan.md §4.
5. Coverage gaps in tasks.md vs spec §4.

Output ONLY strict markdown matching this schema, no preamble, no chat:

# Codex review — <slug>/<artefact>
**Date:** <ISO>
**Verdict:** PASS | FAIL | NEEDS_REWORK
## Findings
- <file:line> — <finding>
## Required fixes (if not PASS)
- <one-liner>
EOF
)
codex exec --sandbox read-only "$PROMPT_TEXT" \
  > "specs/<slug>/reviews/codex-<artefact>-$ISO.md"

Compare verdicts. Read both reports. Print to chat: verifier verdict, codex verdict, points of agreement, divergences.
Lock if both PASS. Update artefact frontmatter Status: locked. Else recommend next step (/clarify, /plan rerun, or /implement on affected tasks).

Output convention

specs/<slug>/reviews/verifier-<artefact>-<ISO>.md — Claude-native challenger
specs/<slug>/reviews/codex-<artefact>-<ISO>.md — adversarial second pair of eyes

Doctrine

Two eyes, not one. Codex must run even if verifier passes. Disagreement is information, not error.
Block locking on FAIL. Either reviewer FAIL ⇒ no Status: locked. User must address.
No edits. Reviews produce verdicts. Fixes happen via /implement re-runs on affected tasks.
Digest freshness check. If supports/digests/latest > 7 days, run ./scripts/sniff-docs.sh first.

Pipeline position

/plan → /review [YOU] → /tasks
/tasks → /review [YOU] → /implement
/implement → /review [YOU] → done