Team Reflective Loop

Team Reflective Loop

Use this skill when a workspace has a .codex-team/ mailbox and the work should proceed as a small reflective team loop rather than a one-shot task.

State Model

Treat the team space as layered state:

• .codex-team/team.md: human-readable brief and slow shared intent.
• .codex-team/team-space.json: workspace identity and protocol metadata.
• .codex-team/tasks/*.json: structured task adapters.
• .codex-team/agents/*.json: supervisor and worker agent registry.
• .codex-team/events/events.jsonl: append-only event history.
• .codex-team/inbox/*.jsonl: per-agent messages.
• .codex-team/failure_bank.json: reusable team error memory.

Read references/team-mailbox-protocol.md before editing mailbox files. Read references/reflective-team-loop.md before deciding the next team action.

Supervisor Loop

For supervisor work:

1. Inspect .codex-team/ and summarize available tasks, agents, blocked work, stale work, and recent events.
2. Read failure_bank.json if it exists and avoid repeating known bad patterns.
3. Pick one bounded next action: assign, request review, add a local patch, mark stale, reassign, or write a handoff.
4. Keep team.md human-readable. Do not use it as the structured source of truth.
5. Prefer append-only events and structured task updates over free-text state.

Worker Loop

For worker work:

1. Read the assigned task JSON.
2. Apply any local_patches before editing.
3. Execute one verifiable slice.
4. Record heartbeat/progress through the CMA UI or by appending structured events only if explicitly asked.
5. Finish with a result envelope: summary, outputs, checks_run, open_risks, and next_request.

Useful Scripts

Run these from the workspace root:

node ~/.codex/skills/team-reflective-loop/scripts/inspect_team_space.js
node ~/.codex/skills/team-reflective-loop/scripts/validate_team_space.js
node ~/.codex/skills/team-reflective-loop/scripts/summarize_team_failures.js

Guardrails

• Do not silently overwrite task owners.
• Do not change board tabs as a side effect of team coordination.
• Do not mark blocked, completed, failed, or review tasks as stale.
• Do not claim success until checks or explicit review evidence exist.
• If a task needs more instruction, add local_patches instead of rewriting the whole brief.