Codex Agent Teams

Codex Agent Teams

Overview

Run one lead agent and multiple specialist teammates to emulate Anthropic Agent Teams patterns in Codex. Keep teammates narrowly scoped, use explicit communication and task-state protocols, and converge through shared quality gates.

This skill mirrors documented Anthropic patterns such as:

• shared task list with pending, in_progress, completed
• teammates claiming independent tasks
• direct teammate messaging and broadcasts
• explicit debate lifecycle (start-debate -> add-position -> decide/apply)
• automatic debate loop orchestration with decision-to-task reflection
• explicit approval checkpoints before implementation
• delegate mode where lead coordinates and does not code

See references/anthropic-pattern-map.md for native-vs-emulated feature mapping.

Routing Gate

Choose the execution mode before dispatch:

• single-agent: one clear change, no coordination benefit.
• single-subagent: moderate scope, one main workstream.
• agent-teams: 2+ independent workstreams, high ambiguity, or strong parallelism benefit.

Promote from single-subagent to agent-teams if scope expands across subsystems or repeated blockers appear.

Startup Sequence

0) Path-safe command setup

Set script paths once so commands work from any directory:

TEAM_SKILL_DIR="${TEAM_SKILL_DIR:-$HOME/.agents/skills/skills_by_openai/codex-agent-teams}"
TEAM_OPS_SCRIPT="$TEAM_SKILL_DIR/scripts/team_ops.py"
TEAM_BRIEF_SCRIPT="$TEAM_SKILL_DIR/scripts/create_team_brief.py"
# Optional explicit override. Leave unset to use auto-discovery.
# TEAM_OPS_ROOT="/absolute/path/to/<workspace>/.codex/teams"

team_ops.py auto-discovers the nearest .codex workspace root when TEAM_OPS_ROOT is not set. Set TEAM_OPS_ROOT only when you intentionally want a non-default storage location.

1) Team bootstrap

Initialize a shared team workspace:

python3 "$TEAM_OPS_SCRIPT" init \
  --team-name "<team-name>" \
  --goal "<goal>" \
  --members "lead,implementer-a,implementer-b,reviewer"

If the team already exists, init is a no-op by default. Use --reset only when you intentionally want to wipe and recreate team state:

python3 "$TEAM_OPS_SCRIPT" init \
  --team-name "<team-name>" \
  --goal "<goal>" \
  --members "lead,implementer-a,implementer-b,reviewer" \
  --reset

If team metadata is partial/corrupted, init automatically recovers state even without --reset.

2) Charter and initial backlog

Create team charter and seed workstreams:

python3 "$TEAM_BRIEF_SCRIPT" \
  --team-name "<team-name>" \
  --goal "<goal>" \
  --workstreams "stream-a,stream-b" \
  --roles "lead,implementer,reviewer,tester" \
  --communication-mode "direct" \
  --delegate-mode \
  --output "<teams-root>/<team-name>/brief.md"

<teams-root> is the same root used by team_ops.py:

• explicit TEAM_OPS_ROOT, if set
• otherwise the auto-discovered <workspace>/.codex/teams

For each work item:

python3 "$TEAM_OPS_SCRIPT" add-task \
  --team-name "<team-name>" \
  --title "Implement X" \
  --owner "unassigned" \
  --status "pending" \
  --depends-on ""

Role prompt hand-off (manual, no auto-generator):

• Build per-role prompts from references/prompt-templates.md by filling placeholders from brief.md, task ownership, and current debate/task state.
• team_ops.py manages state/logs only; it does not synthesize role prompts.
• For traceability, include role id, owned scope, required outputs, and exact team_ops.py commands each sub-agent should run.

3) Approval checkpoint

Before coding, present team plan and wait for explicit user approval. For strict planning phases, pair with:

• superpowers:brainstorming
• superpowers:writing-plans

4) Optional delegate mode

If delegate mode is on, lead does orchestration only:

• assign and re-balance work
• resolve cross-stream conflicts
• enforce convergence gates
• avoid direct code edits unless user asks

Communication Protocol

Use shared mailbox commands to emulate direct teammate communication:

Send direct message:

python3 "$TEAM_OPS_SCRIPT" message \
  --team-name "<team-name>" \
  --from "implementer-a" \
  --to "reviewer" \
  --body "Please review task-3 patch and edge-case notes."

Broadcast update:

python3 "$TEAM_OPS_SCRIPT" broadcast \
  --team-name "<team-name>" \
  --from "lead" \
  --body "Task dependency changed: api-schema now blocks frontend-sync."

Read inbox:

python3 "$TEAM_OPS_SCRIPT" inbox --team-name "<team-name>" --member "reviewer"

Rules:

• Every teammate reports decisions and blockers through messages.
• Critical dependency changes are broadcast to all members.
• Lead records final decisions in brief and task board.
• For visible team behavior, prefer multiple message rounds (direct messages and inbox reads) instead of a single end-state broadcast.

Sub-agent guardrails:

• --team-name must be a single identifier (no /, \\, ., .., whitespace, or control characters) for both team_ops.py and create_team_brief.py.
• --from, --to, --member, --members, --decider must refer to registered team members.
• --owner accepts either a registered team member or unassigned.
• create_team_brief.py --roles defines descriptive role labels only; command flags still require actual team member IDs from team.json.
• --owner-map owners must be registered team members or unassigned (one mapping per option; duplicates are rejected). Owners are not limited to debate participants.
• start-debate and new-debate orchestrate-debate default --decider to lead; if team has no lead, the command falls back to the first debate member with a warning (explicit --decider is still recommended).
• New debate creation requires at least two unique --options and at least two unique registered --members; otherwise commands fail fast with explicit input guidance.
• add-position --member must be one of the debate's registered members (not just any team member).
• Unknown member/task/debate ids include contextual guidance (registered members or available ids, plus typo suggestions when detectable), including --notify-from validation errors.
• If a team is missing under the current root and --team-root is not explicitly set, commands auto-switch to a uniquely discovered ancestor .codex/teams root with a warning.
• If --team-root is explicitly set (or TEAM_OPS_ROOT is set), commands do not auto-switch roots; they fail under the specified root for deterministic behavior.
• --team-root (or TEAM_OPS_ROOT) must resolve to a directory path (not a file path).
• Mutating commands serialize writes with a per-team lock at the resolved team root (after implicit auto-switch when applicable) to prevent concurrent state clobbering; if lock wait times out, retry or tune TEAM_OPS_LOCK_WAIT_SECONDS.
• Stale lock eviction checks recorded PID liveness and lock identity; active or newly replaced locks are not reclaimed solely by age.
• Commands fail fast on non-member identities to prevent mailbox/task-board drift.

Task-State Protocol

State model:

• pending: not started or waiting on dependency
• in_progress: claimed by one teammate
• completed: done and verified

Claim task:

python3 "$TEAM_OPS_SCRIPT" claim \
  --team-name "<team-name>" \
  --task-id "task-2" \
  --member "implementer-b"

Update task status:

python3 "$TEAM_OPS_SCRIPT" update-task \
  --team-name "<team-name>" \
  --task-id "task-2" \
  --status "completed" \
  --note "Unit and integration checks passed."

List task board:

python3 "$TEAM_OPS_SCRIPT" list-tasks --team-name "<team-name>"

Debate Protocol (Conflict to Decision)

When two or more approaches conflict, run a formal debate and persist the outcome. Debates are stored in <teams-root>/<team-name>/debates.json.

Storage boundaries:

• debates.json stores structured debate state only (topic, options, members, positions, decision, applied state).
• messages.jsonl stores communication events only (direct/broadcast messages).
• start-debate --notify fans out one direct message per debate member in messages.jsonl (not a single broadcast record).
• Sending a message does not create a debate position; only add-position appends to debates.json.

Start a debate:

python3 "$TEAM_OPS_SCRIPT" start-debate \
  --team-name "<team-name>" \
  --topic "Choose retry strategy for API sync" \
  --task-id "task-2" \
  --options "fixed-backoff,exponential-backoff" \
  --members "lead,implementer-a,implementer-b,reviewer" \
  --decider "lead" \
  --notify

Each member submits a position:

python3 "$TEAM_OPS_SCRIPT" add-position \
  --team-name "<team-name>" \
  --debate-id "debate-1" \
  --member "implementer-a" \
  --option "exponential-backoff" \
  --confidence 0.8 \
  --rationale "Lower p95 under burst failures."

--confidence must be a finite numeric value in [0, 1].

Lead decides and applies to the linked task:

python3 "$TEAM_OPS_SCRIPT" decide-debate \
  --team-name "<team-name>" \
  --debate-id "debate-1" \
  --rationale "Best reliability/cost tradeoff from submitted evidence." \
  --require-all-positions \
  --apply \
  --status-on-apply "in_progress" \
  --owner-map "fixed-backoff:implementer-a,exponential-backoff:implementer-b"

Review debate state:

python3 "$TEAM_OPS_SCRIPT" list-debates --team-name "<team-name>"
python3 "$TEAM_OPS_SCRIPT" show-debate --team-name "<team-name>" --debate-id "debate-1"

Interaction fidelity recommendation:

• When you want explicit teammate interaction traces, it is desirable to run multiple rounds of message and add-position, while keeping decide-debate as a single final step.
• Suggested pattern:

1. start-debate with --notify to open the decision and notify participants.
2. Each member reads inbox and sends at least one direct message to challenge or clarify assumptions.
3. Each member submits add-position while debate status is open (revisions are allowed only before decision).
4. Lead sends a final synthesis message or broadcast, then runs decide-debate.

• This produces an auditable collaboration trail in messages.jsonl and debates.json rather than a one-shot decision record.

Decision semantics:

• decide-debate defaults the decider to the debate's configured decider when --decider is omitted.
• --rationale is required when recording a new decision.
• For an already decided debate, decide-debate --apply reflects the existing decision to task state; it does not allow changing the selected option.
• For an already applied debate, decide-debate exits as no-op before --owner-map parsing/validation.
• --owner-map is validated only when applying to a linked task (task_id present); debates without linked tasks ignore owner mapping.
• To change direction after a decision, open a new debate linked to the blocked task.

Automatic Orchestration Loop

Use one command to run the loop:

1. create/load debate
2. remind missing members
3. auto-decide when all positions are present (weighted confidence)
4. reflect decision into linked task (status, optional owner mapping, note, broadcast)

python3 "$TEAM_OPS_SCRIPT" orchestrate-debate \
  --team-name "<team-name>" \
  --topic "Choose cache invalidation strategy" \
  --task-id "task-4" \
  --options "ttl-only,event-driven" \
  --members "lead,implementer-a,implementer-b,reviewer" \
  --decider "lead" \
  --send-reminders \
  --status-on-apply "in_progress" \
  --owner-map "ttl-only:implementer-a,event-driven:implementer-b"

Re-run orchestrate-debate until applied. It is idempotent once a decision is applied. For an already applied debate, orchestrate-debate exits as no-op before --owner-map parsing/validation. For debates that are still waiting on positions, orchestrate-debate defers --owner-map validation until the apply step. If an existing decision payload is corrupted (for example, decision option not in debate options), orchestration fails fast instead of applying invalid state.

Reminder/broadcast sender behavior:

• --notify-from is validated only when reminders or apply broadcasts are actually emitted.
• If --notify-from is left as default lead and lead is not in team members, sender falls back to the active decider.
• The same fallback rule applies to start-debate --notify.

Monitoring (Optional, Default OFF)

team_ops.py supports opt-in monitor logging for command-level observability. Monitoring is disabled by default and no monitor file is created unless explicitly enabled.

Enable monitoring for a command:

python3 "$TEAM_OPS_SCRIPT" orchestrate-debate \
  --team-name "<team-name>" \
  --debate-id "debate-1" \
  --monitoring

Override monitor file path:

python3 "$TEAM_OPS_SCRIPT" update-task \
  --team-name "<team-name>" \
  --task-id "task-1" \
  --status "completed" \
  --monitoring \
  --monitor-log-file "/tmp/team-monitor.jsonl"

Environment alternatives:

• TEAM_OPS_MONITORING=1
• TEAM_OPS_MONITOR_LOG_FILE=/path/to/monitor.jsonl
• TEAM_OPS_ROOT=/path/to/.codex/teams
• TEAM_OPS_LOCK_WAIT_SECONDS=30 (optional write-lock wait override)
• TEAM_OPS_LOCK_STALE_SECONDS=600 (optional stale-lock eviction window)

Generate monitor summary report:

python3 "$TEAM_OPS_SCRIPT" monitor-report \
  --team-name "<team-name>" \
  --output "<teams-root>/<team-name>/monitor-report.json"

monitor-report is team-filtered by team_name, and reports:

• invalid_event_lines: malformed lines skipped (including missing/invalid team_name)
• other_team_event_lines: schema-valid lines skipped because they belong to another team
• ISO timestamps without timezone offset are interpreted as UTC when computing reflection latency.
• at accepts common ISO-8601 UTC/offset forms (for example ...Z, ...+00, ...+00:00, ...+0000).
• Missing required event fields (for example at, event_type, command, actor, entity_type, entity_id) or invalid at timestamp formats are treated as invalid monitor lines.

Monitoring event schema (monitor.jsonl, one JSON object per line):

For debate.applied, command is logged as apply-decision (internal apply step used by both decide-debate --apply and orchestrate-debate).

Example line:

{"at":"2026-02-16T06:00:00+00:00","event_type":"debate.applied","command":"apply-decision","team_name":"example-team","actor":"lead","entity_type":"debate","entity_id":"debate-1","before":{"task":{"owner":"unassigned","status":"pending"}},"after":{"task":{"owner":"implementer-a","status":"in_progress"}},"metadata":{"task_id":"task-1","option":"session","status":"in_progress","sender":"lead"},"correlation_id":"2e5db8f0d25d4f08b6f2f4f13a4ed8ad"}

Superpowers Integration

Map workstreams to one control skill:

• plan refinement and ambiguity reduction: superpowers:brainstorming then superpowers:writing-plans
• independent implementation in this session: superpowers:subagent-driven-development
• multiple unrelated failures or investigations: superpowers:dispatching-parallel-agents
• execution from approved written plan in batches: superpowers:executing-plans
• isolated workspace setup: superpowers:using-git-worktrees
• pre-completion validation: superpowers:verification-before-completion
• final quality check: superpowers:requesting-code-review
• closeout: superpowers:finishing-a-development-branch

Convergence Gates

Before merging teammate outputs:

1. Spec compliance gate
2. Code quality gate
3. Verification gate (superpowers:verification-before-completion)
4. Final review for major diffs (superpowers:requesting-code-review)
5. Branch closeout (superpowers:finishing-a-development-branch)

Operating Rules

• Keep teammates specialized and replaceable.
• Prefer teams of 2 to 5 members.
• Keep one source of truth for status in team task board.
• Do not mark task complete without command output.
• Escalate unresolved blockers after two failed attempts.
• If two solutions conflict, start a debate and link it to the blocked task.
• Do not close debates without recorded rationale and selected option.
• Reflect debate outcomes into task state before resuming implementation.

Limitations

This skill emulates Agent Teams patterns in Codex, but some native Anthropic UX features may not exist in every runtime. When unavailable, fall back to:

• explicit message commands via scripts/team_ops.py
• shared files for task board and decisions
• lead-mediated routing while preserving protocol semantics

Resources

• scripts/create_team_brief.py: generate a reusable team charter and workstream summary template (canonical task board state lives in scripts/team_ops.py).
• scripts/team_ops.py: manage team members, task states, mailbox, debates, and auto-orchestration.
• references/anthropic-pattern-map.md: what is mirrored from Anthropic docs.
• references/team-topologies.md: choose collaboration pattern by coupling and risk profile.
• references/prompt-templates.md: copy/paste lead, specialist, reviewer, and challenger prompts.