cxb-review

Dual assessment for step-level and task-level review.

What It Does

cxb-review provides a second layer of quality control by combining Claude's local assessment with a cross-review from another provider, usually Codex.

This matters because a single model can miss:

• hidden regressions
• requirement drift
• weak tests
• incomplete acceptance criteria coverage

Review Modes

Mode	When it runs	What it checks
step	After one execution step	Done conditions for that step
task	After all steps are complete	Full acceptance criteria and overall quality

Decision Flow

1. Claude Assessment

Claude evaluates the work first and assigns one of:

• PASS
• FIX
• UNCERTAIN

2. Cross-Review Assessment

The cross-review provider answers:

1. Does it agree with Claude's verdict?
2. What did Claude miss?
3. Should the work pass or be fixed?

3. Final Decision

Claude	Cross-review	Result
PASS	PASS	PASS
PASS	FIX	FIX
FIX	PASS	FIX
FIX	FIX	FIX
UNCERTAIN	any	Claude makes the final call

The system intentionally biases toward fixing when there is disagreement.

Step Mode Checklist

For a single execution step, review should answer:

• Were the done conditions met exactly?
• Did the code change stay within scope?
• Was any obvious regression introduced?
• Were tests run if this step required them?

Task Mode Checklist

For the whole task, review should answer:

• Are all acceptance criteria satisfied?
• Are docs or config updates missing?
• Are there remaining correctness or maintainability problems?
• Are follow-up tasks needed rather than silent compromises?

Example Review Output

{
  "mode": "step",
  "target": "Add audit logging to admin mutations",
  "verdict": "FIX",
  "claudeAssessment": {
    "verdict": "PASS",
    "reason": "Done conditions appear satisfied"
  },
  "crossAssessment": {
    "verdict": "FIX",
    "agreedWithClaude": false,
    "missedIssues": [
      "DELETE mutations are not logged",
      "No test covers actor attribution"
    ],
    "fixItems": [
      "Log destructive admin actions",
      "Add regression test for actor_id capture"
    ]
  },
  "finalDecision": {
    "verdict": "FIX",
    "reason": "Cross-review found requirement gaps"
  }
}

Best Practices

• Keep review binary: pass or fix. Avoid soft "mostly good" conclusions.
• Require specific fix items, not vague advice.
• Use task-mode review before reporting completion on anything user-visible or risky.
• If Claude and Codex disagree, inspect the disagreement. That is often where the real issue is.

Good Requests To Trigger cxb-review

• "Run a step-level review against the done conditions."
• "Review this completed task against the original acceptance criteria."
• "Have Codex challenge the implementation and only pass if tests and docs are complete."