Codex and Claude collaboration

Operating model

One agent writes; the other reviews. Roles can rotate between tasks, but ownership cannot be ambiguous within a task.

Handoff format

Task: TASK-###
Role completed: implementer | reviewer
Commit/diff: <identifier>
Writable scope used: <paths>
Acceptance criteria status: <list>
Checks run: <commands and outcomes>
Open findings: <P0-P3 list>
Known limitations: <list>
Next owner/action: <one concrete action>

Conflict handling

Product requirements and human decisions win.
AGENTS.md wins over agent-specific files.
Task acceptance criteria win over implementation preference.
Automated evidence wins over an unsupported assertion.
Security-sensitive ambiguity is escalated to the human owner.

Review independence

The reviewer must inspect the actual diff and tests. Do not provide the implementer's reasoning as a substitute for evidence; it can anchor the reviewer and hide alternative failure modes.

Serial development

Implementation is serial: one writer owns the working tree at a time, and both agents append completion and handoff records to docs/development-log.md. The current ownership map and delivery sequence are recorded in docs/workstreams.md.

1.2 KiB Raw Blame History