/review — AI critic, then a human survey

The review-needed/ lane held finished work for human eyes. /review is how those eyes work: one plan, an AI pass, then Approve / Needs Work / Skip.

The missing UX

Moving a plan into review-needed/ was already doctrine. What was missing was a human-facing session that:

1. Checks out feature/<slug> without stomping a dirty tree
2. Runs a fresh-context critic against the plan and the branch diff
3. Asks a clear survey instead of hoping free-text “lgtm” means what you think
4. Turns “needs work” into actionable notes and a return to the ready queue (bugs/ or features/), not a silent re-claim to in-progress/

What /review does

One plan per invocation. Bare /review picks by Priority → Value → oldest mtime; other queued plans get a one-line mention only. Re-run for the next item.

Pipeline (fixed):

select → checkout → evidence + optional launch
       → AI code review (fresh context)
       → present package
       → survey → follow-ups → lane move

Survey	Move
Approve	→ completed/ (override required if the critic said REVISE/ESCALATE)
Needs Work	→ bugs/ or features/ (same inference as /draft --promote)
Skip	stay in review-needed/

AI is advisory. The human survey is authoritative. Agents still must not complete from review-needed/ outside a confirmed Approve.

Where it ships

• Claude: .claude/commands/review.md
• Grok: .grok/skills/review/SKILL.md
• Docs: Skills → /review
• Scaffolded into projects with the other dual-use skills

Pair with the earlier review-needed lane post: lane first, skill second — “done” is no longer the same event as “completed.”