cross-reviewFor Claude Code and CodexMITNode.js 22+
A second model checks your agent's work.
Your agent writes the plan and the code. cross-review hands them to a second model, which reads the repository, runs the tests and writes down what is wrong. The agent fixes each issue or answers it with evidence, round after round, until the reviewer approves.
npx @tlelabs/cross-reviewInstalls six skills for Claude Code and Codex, and the cross-review MCP server they use.
ISSUE-1 high A quoted field with a line break is split into two rows. src/export/csv.ts:42
ISSUE-2 medium JSON export drops null fields.
ISSUE-1 fixed: fields are quoted by the RFC 4180 rules, with a regression test. ISSUE-2 disputed: the plan's phase 2 drops null on purpose, to keep the files small.
Ran npm test: 214 passed. The plan confirms ISSUE-2 is intended; withdrawn.
VERDICT: APPROVE
How a review goes
Two models, one thread, until they agree.
The review runs in the background while you keep working. The agent you talk to drives it; the reviewer only reads, runs and reports.
- 01
The agent asks
After a plan or a change, the skill sends the reviewer what to check: the plan file, the diff, the question.
- 02
The reviewer checks
It reads the code, runs tests and commands, and writes each finding as an issue with a place and a severity.
- 03
The agent answers
Every issue is fixed or disputed with evidence: a file and line, a test result, your earlier decision.
- 04
Approve or ask you
Rounds repeat until the reviewer approves. If they cannot agree, the agent brings the open points to you.
Skills
Six skills: plan, build, review.
Run them by name: /review-impl in Claude Code, $review-impl in Codex. With--auto, Claude Code reviews every plan and change on its own.
| Skill | Use it | What happens |
|---|---|---|
| Plan and build | ||
| loop-plan | For large work: several features, many files, more than one session. | Writes a phased plan grounded in the code, with tests per phase and review checkpoints, then has it reviewed. |
| loop-implement | Once the plan is approved. | Builds it phase by phase with tests first, ticks each criterion it meets, commits, and runs a review at each checkpoint. |
| Review | ||
| review-plan | Before you write code. | Debates the plan: what will not work, missing steps, claims that do not hold. |
| review-impl | Before you commit, push or merge. | Reviews uncommitted changes, unpushed commits or a branch against the plan, and the agent fixes what is real. |
| review-think | When a decision needs a second opinion. | Both models reason on their own, then compare: what changed their minds, and with what evidence. |
| review-pr | On someone else's pull request. | Reviews it as a peer: correctness, the description, and whether it is ready to merge. |
Dashboard
Watch every review, round by round.
A local page lists the reviews and plans in your projects. A running review updates live: what the reviewer is reading or running right now, each round's issues, and how the agent answered.
$ cross-review dashboard
You decide
The reviewer advises. You have the last word.
- Disputes need evidence. The agent rejects a finding only with a file, a test or a decision you made.
- Stalemates come to you. When the two models cannot agree, the open points are put to you as a choice.
- Your instruction wins. Skip a review, accept a result, change the model: the agent follows what you say last.
Local
Everything stays on your machine.
- Sessions are files. Each review is kept in your project's
.cross-review/, which git ignores. - The dashboard is private. It listens on 127.0.0.1 only, with a token per run.
- The reviewer only reports. It may run tests and commands to check its claims; its instructions forbid changing your files.
Get started
Install once, for both agents.
npx @tlelabs/cross-review