You MUST read AGENTS.md before any tool call, file write, or gate decision. If it is missing,
stop before implementation or gate decisions unless the active issue is specifically restoring it.
You MUST read docs/agent-workflow.md before starting issue work.
Codex is a single-agent executor inside this project's multi-role, phase-driven workflow. Codex runs one issue through explicit role passes and records machine-checkable evidence instead of relying on implicit persona switching alone.
- Run the issue via
agents/workflows/orchestrate/SKILL.md - Execute one formal phase at a time
- Read the previous role-pass before starting the next one
- Write a new role-pass artifact after every completed phase
- Record the actual executor name in every artifact and the model / runtime when known; never default to another adapter name
- Keep the workflow-status comment aligned with the latest phase state
- Bounded and standard work: self-review is allowed, but it must be explicit and evidence-backed
- High-assurance work: stop and request human security/acceptance review. This review happens at the PR stage — open the PR first, then request review; implementation commits, pushes, and PR creation are never blocked on it (
docs/agent-workflow.md§8) - Review roles are read-only unless the request explicitly returns to implementation
- Validate any subagent output before incorporating it into workflow artifacts or commits
- Use multi-agent delegation only for broad discovery, advisory review, or real async support
- Record follow-up findings as issues, not TODOs
The default headless CLI is codex exec. Set CODEX_CLI to override it when needed.
- If Codex is unavailable for implementation, route to Claude, then Agy
- If Codex is unavailable for orchestration or review support, route to Claude or Agy
- If no qualified reviewer is available for a required human gate, stop for human review