codex-multi-agents-flowlisted

# Codex multi-agent flow Deliver one scoped story with an implementer session and an independent reviewer session. Prefer Codex-native worktrees, review commands, JSONL traces, and repository instructions. Treat another provider such as Claude Code as an optional replacement for either role, not as a requirement. ## Guardrails 1. Keep one story on one branch and one worktree. Only the implementer may edit it. 2. Preserve the user's current branch and uncommitted changes. Never reset, clean, checkout, rebase, push, open a PR, or merge unless the user explicitly places that action in scope. 3. Materialize the story in `specs/<ID>.md` before implementation. Use a connected source-of-truth tool only when needed, fetch the narrowest sufficient context, and retain source links or IDs in the local spec. 4. Start review from the branch diff, then inspect related callers, tests, schemas, configuration, and invariants when needed to validate a finding. Do not impose a diff-only context limit. 5. Require each finding to include severity, file and line, failure scenario, and a minimal correction. Report no finding when evidence is insufficient. 6. Allow at most one fix pass and one optional re-review. Escalate unresolved design decisions to the user instead of starting agent-to-agent loops. 7. Define completion with commands and artifacts. Agent prose is not verification. 8. Keep PR creation and merge as explicit human gates. ## Prepare the story Read all app