Use the other coding harness as the independent reviewer, then report in the compact review verdict shape.
Workflow
- Define the review target from the user's request: uncommitted diff, branch vs base, commit, PR, or named paths. If unspecified, review the current uncommitted diff.
- Detect the active harness:
- In Codex/OpenAI, invoke Claude Code for the review.
- In Claude/Anthropic, invoke Codex for the review.
- Run the preflight checks for the opposite harness.
- Delegate with the shared prompt below, filled with the target.
- Preserve the other harness's findings; note any local confirmation or dispute.
- Stop after one other-model review unless the user explicitly asks for an iterative loop.
For CLI flag details, fallback commands, and hosted-review cautions, read references/headless-modes.md.
Preflight
Check only the opposite harness:
# Codex active -> Claude reviewer
command -v claude
claude --version
claude auth status
# Claude active -> Codex reviewer
command -v codex
codex --version
codex login status
Report only installed/authenticated status. Keep account details, tokens, and config contents out of the report.
Shared Prompt
Use /review or $review if available. Otherwise follow the uinaf review contract:
findings first, then verdict exactly one of ship it / needs review / blocked,
then evidence, unverified, next, and optional notes.
Target: <uncommitted diff | branch vs base | commit | PR | paths>
Review read-only. Leave files, staging, commits, and fixes untouched.
Prioritize bugs, regressions, missing tests, silent failures, and contract drift. Use concise evidence with file/line references and targeted command results.
Codex Active
Use Claude Code print mode:
claude -p \
--permission-mode dontAsk \
--allowedTools "Read,Glob,Grep,LS,Bash(git *),Bash(rg *),Bash(jq *),Bash(npm *),Bash(pnpm *)" \
"<review prompt>"
Use hosted claude ultrareview only when the user explicitly asks and accepts the cost.
Claude Active
Use Codex exec review:
codex exec review --uncommitted --ephemeral - <<'PROMPT'
<shared prompt with Target set to current uncommitted diff>
PROMPT
Variants: branch review uses --base <base-ref>, commit review uses --commit <sha>, and custom scope belongs in the prompt.
Output
Match review:
- findings: none
- verdict: ship it
- evidence: Claude review covered the uncommitted diff and found no material issues
- unverified: runtime smoke not rerun
- next: verify
If the opposite harness cannot run:
- findings: none
- verdict: blocked
- evidence: Claude Code is not installed or authenticated, so no independent cross-harness review ran
- unverified: review target not inspected by the opposite harness
- next: review