speckit.boundedrelay.implementation-review

Dual Implementation Review

Review the phase named in $ARGUMENTS against its pending revision-sealed evidence. This command is read-only except for the matching run-local review JSON.

1. Read the stage in $ARGUMENTS. For stage=claude, read the constitution, active spec, plan, tasks, repository rules, real diff, and checks. Complete only the Claude host review, set the evidence state to claude-frozen, keep Codex evidence and the combined verdict null, atomically replace the file, and stop. Preserve implementation check receipts. For convergence changes, run the relevant checks and add redacted typed receipts plus their canonical digest before freezing. Every receipt's tested Git tree must equal the sealed convergence revision tree; raw output and environment values never enter the evidence file. The following verifier stamps an engine-owned reviewId bound to the run, nonce, revision, source evidence, and check digest. Preserve that exact ID when submitting hostReview to BoundedRelay.
2. For stage=codex, require that the separate verifier already accepted the frozen Claude state. Do not edit or regenerate the host review and do not put its findings in the Codex prompt or focus.
3. Start a fresh codex_worker_sdd_review with the matching implementation or convergence phase, mode: strict, exact sealed artifact paths, the recorded comparison base revision, expected full Git revision, every sealed artifact path including a configured project profile, and frozen host evidence. Use the exact prepared codexReviewPolicy (including a profiled cross-review model and effort), and persist the job's observed model and reasoning effort. Generic codex_worker_analyze cannot satisfy this gate. Keep Claude findings out of focus and the Codex prompt. Use an allowlisted model only, poll with afterRevision, and retrieve the final result once.
4. Recompute workspace state. Evidence is invalid if HEAD, changed-file scope, artifacts, source execution/review digest, check receipts, or the working-tree seal changed.
5. Record severity-ranked findings with exact paths, verification observed, and residual risk. High and Critical findings are always blocking and require a new revision plus two fresh reviews. A missing or failed reviewer is not replaced by self-review.
6. Persist the returned strict seal and gate. Set the combined verdict to approved only when both reviews approve and the returned gate is passed: true, status: ready for that seal. Atomically replace the evidence. Do not fix findings, edit tasks, apply patches, commit, push, or deploy.

$ARGUMENTS