Codex-Claude Duo (1.x legacy)

This page preserves 1.x behaviour and commands, not the native 2.x setup. Use the 2.x guide for current installation and usage.

Use duo mode when Codex and Claude should provide complementary review without requiring a third runtime. Codex is strongest for implementation discipline, command evidence, artifact paths, and reproducibility. Claude is strongest for scholarly argument, reviewer empathy, manuscript structure, and assumption pressure.

Common Patterns

Codex-primary:

python3 -m bridges.orchestrator task-run \
  --task-id I7 \
  --paper-type methods \
  --topic policy-effects \
  --cwd . \
  --execution-mode duo \
  --controller codex \
  --primary codex \
  --reviewer claude \
  --mcp-strict \
  --skills-strict

Claude-primary:

python3 -m bridges.orchestrator task-run \
  --task-id H3 \
  --paper-type empirical \
  --topic ai-in-education \
  --cwd . \
  --execution-mode duo \
  --controller claude \
  --primary claude \
  --reviewer codex \
  --mcp-strict \
  --skills-strict

Review Contract

Each duo run should produce or preserve:

• primary artifact under RESEARCH/[topic]/
• reviewer notes with explicit accept/block/revise status
• templates/duo-review-report.md when a human-readable review artifact is needed
• templates/disagreement-matrix.md when Codex and Claude conflict on evidence, method, implementation, or acceptance

Disagreement Matrix

Case	Codex emphasis	Claude emphasis	Decision rule
Citation or claim support	Check artifact paths, ledgers, and reproducibility of evidence capture.	Check claim scope, venue expectations, and reader interpretation.	Narrow or mark the claim as a gap unless both evidence traceability and rhetorical scope are acceptable.
Method design	Check implementation feasibility, data requirements, and validation commands.	Check construct validity, alternative explanations, and reviewer objections.	Block until both executable method and defensible research design are present.
Code change	Check tests, diff scope, and rollback path.	Check whether the code still answers the research question and explains limitations.	Accept only when command evidence and research rationale agree.
Manuscript revision	Check required outputs, references, and contract conformance.	Check coherence, tone, reviewer empathy, and contribution clarity.	Revise until the artifact is both contract-valid and publication-facing.
Residual uncertainty	Preserve unresolved technical risk.	Preserve unresolved scholarly risk.	Escalate to --triad or human adjudication if the risk changes conclusions or downstream decisions.

Do not average conflicting judgments. Record the concrete issue, cite the artifact or evidence reference, assign risk level, and choose accept, revise, block, or escalate.