Review Spec

Assess a thread-root spec.md as a downstream handoff, strictly read-only. The one question you answer: could another agent plan and implement from this spec alone, without any hidden conversational context? You read the spec as that downstream agent would, judge it against the axes below, and record any findings for later attention. You never touch the spec or any other artifact, and you produce no durable report — a clean spec earns a chat judgment and nothing on disk.

This is a quality-of-handoff and planning-readiness review; your concern is whether the document, as written, is fit for someone else to act on safely.

Procedure

1. Resolve the thread. Work inside one thread root at docs/threads/<YYMMDDHHMMSSZ-slug>/. If cwd already sits inside a thread root, that is the thread. Two situations make a findings bundle physically impossible — .pending-reviews/ would live inside the very thread that failed to resolve — so in both, refuse in chat, write nothing, and end with Outcome: REFUSED — <reason>: no thread exists yet, or several thread roots exist and which is active is ambiguous (never silently pick the most recent stamp).

2. Read the spec read-only. The target is the thread-root spec.md. Read it end to end at least once, as a downstream planner with no memory of the conversation that produced it. If no spec.md exists at the thread root, tell the user there is nothing to review, write nothing, and end with Outcome: REFUSED — no spec.md to review.

3. Judge against the readiness axes. Assess the spec on each axis below (## What you judge). For every real weakness, form a finding: what is wrong, where in the spec it shows, why it would leave a downstream agent guessing or blocked, and a severity — blocker (planning cannot proceed safely), issue (a real gap that will cause rework or a wrong guess), or nit (soft or imprecise, but survivable). Tether every finding to downstream impact: "this is vague" is not a finding; "this is vague, so a planner must guess whether X means A or B" is.

4. Decide the outcome. If the spec passes the bar — a downstream agent could plan and implement from it without hidden context — the review is clean. Otherwise you hold one or more findings to record.

5. Report. A clean review returns a concise readiness judgment in chat and writes no file, ending Outcome: DONE — <the readiness judgment>. A review with findings emits exactly one bundle (## Recording findings), reports its path, and ends Outcome: DONE — findings at <bundle path>. No preamble, no closing remark.

What you judge