AGENT SKILL

cross-review

Open a visible T3 review thread where a second model (Astra Extra High for a Claude author, Opus for a Codex author) reviews a plan or change, then go back and forth in that thread until both accept the same version. The owner can join the thread and steer. Use only when the operator asks for it (/cross-review or an explicit request); never start one on your own.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 0
Forks
0

01 前言 Front Matter

name
cross-review
description
Open a visible T3 review thread where a second model (Astra Extra High for a Claude author, Opus for a Codex author) reviews a plan or change, then go back and forth in that thread until both accept the same version. The owner can join the thread and steer. Use only when the operator asks for it (/cross-review or an explicit request); never start one on your own.

03 SKILL.md 正文

Cross review

The author keeps writing; the reviewer only proposes. The conversation lives in one ordinary T3 thread the owner can open, read and write in: the owner asked for one continuous visible thread, so this is the explicit exception to fresh review rounds. It does not replace the Sol and Codex reviews the global review budget requires for a PR: one Sol review of the finished PR (none for docs-only or small low-risk PRs) and Codex asked at most twice. The reviewer model is the owner's choice: Astra Extra High for a Claude author, Opus for a Codex author; the owner may name another.

Start

  1. Freeze the artifact: a file under $HOME/.development-system/private/cross-review/<id>/ (plan, list or contract) or a PR head SHA. Compute its SHA-256.
  2. Record <id>.json there: author thread id, artifact path and hash, round 0.
  3. Check t3_thread_list for an existing review thread for this id before launching; a launch has no retry key.
  4. t3_thread_launch in the same project and workspace as the author (workspaceStrategy of the author's checkout; never a new worktree), with:
    • model: codex / gpt-6-astra, reasoningEffort: xhigh for a Claude author; claudeAgent / claude-opus-5-5 for a Codex author;
    • first message: the owner's goal in your own words, the artifact path and hash, the evidence it rests on, "read only: do not edit, commit, switch branches or start servers", and the answer format below.
  5. Save the returned thread id and run id; link the thread to the owner as [title](t3-thread://v1/<threadId>).

Reviewer answer format: restate the owner's goal, then numbered findings (R<round>-H<n>) each with severity (blocking, important, minor), evidence and a concrete change; end with exactly VEREDICTO: SIN BLOQUEANTES or VEREDICTO: CON BLOQUEANTES, and when there are no blockers, whether it accepts the artifact with that hash.

Each round

  1. Wait for the recorded run with t3_thread_wait and that runId. Timeout, interruption or failure leave the round pending; never read it as acceptance.
  2. Read the whole answer with t3_thread_read (page with afterPosition, recover truncated text with itemId/textOffset).
  3. If the owner wrote in the thread since the last round, the owner's message wins: reconcile it before answering the reviewer.
  4. For every finding: fix the artifact, or reject it with evidence. Save the new version, compute its hash, add a response table at the end.
  5. Send the response with t3_thread_send, mode queue, a stable clientRequestId (<id>-r<round>), the new path and hash, and the request to verify each fix and look for regressions. Record the new messageId/runId.

Close

  • Agreed: the reviewer says SIN BLOQUEANTES and accepts the current hash. The author uses exactly that version.
  • Disagreement: five rounds without agreement, or a finding the author rejects twice. Report both positions to the owner; this is never approval.

Report the rounds, the accepted hash and the thread link. The review validates a design; the work it describes is still pending until done and verified.

04 同仓库更多 Skills