Team

Team Skill

Spawn N coordinated agents working on a shared task list using Codex's native team tools. Replaces the legacy /swarm skill (SQLite-based) with built-in team management, inter-agent messaging, and task dependencies -- no external dependencies required.

Usage

/oh-my-codex:team N:agent-type "task description"
/oh-my-codex:team "task description"
/oh-my-codex:team ralph "task description"

Parameters

• N - Number of teammate agents (1-20). Optional; defaults to auto-sizing based on task decomposition.
• agent-type - OMC agent to spawn for the team-exec stage (e.g., executor, build-fixer, designer). Optional; defaults to stage-aware routing (see Stage Agent Routing below).
• task - High-level task to decompose and distribute among teammates
• ralph - Optional modifier. When present, wraps the team pipeline in Ralph's persistence loop (retry on failure, architect verification before completion). See Team + Ralph Composition below.

Examples

/team 5:executor "fix all TypeScript errors across the project"
/team 3:build-fixer "fix build errors in src/"
/team 4:designer "implement responsive layouts for all page components"
/team "refactor the auth module with security review"
/team ralph "build a complete REST API for user management"

Architecture

User: "/team 3:executor fix all TypeScript errors"
              |
              v
      [TEAM ORCHESTRATOR (Lead)]
              |
              +-- TeamCreate("fix-ts-errors")
              |       -> lead becomes team-lead@fix-ts-errors
              |
              +-- Analyze & decompose task into subtasks
              |       -> explore/architect produces subtask list
              |
              +-- TaskCreate x N (one per subtask)
              |       -> tasks #1, #2, #3 with dependencies
              |
              +-- TaskUpdate x N (pre-assign owners)
              |       -> task #1 owner=worker-1, etc.
              |
              +-- Task(team_name="fix-ts-errors", name="worker-1") x 3
              |       -> spawns teammates into the team
              |
              +-- Monitor loop
              |       <- SendMessage from teammates (auto-delivered)
              |       -> TaskList polling for progress
              |       -> SendMessage to unblock/coordinate
              |
              +-- Completion
                      -> SendMessage(shutdown_request) to each teammate
                      <- SendMessage(shutdown_response, approve: true)
                      -> TeamDelete("fix-ts-errors")
                      -> rm .omc/state/team-state.json

Storage layout (managed by Codex):

~/.codex/
  teams/fix-ts-errors/
    config.json          # Team metadata + members array
  tasks/fix-ts-errors/
    .lock                # File lock for concurrent access
    1.json               # Subtask #1
    2.json               # Subtask #2 (may be internal)
    3.json               # Subtask #3
    ...

Staged Pipeline (Canonical Team Runtime)

Team execution follows a staged pipeline:

team-plan -> team-prd -> team-exec -> team-verify -> team-fix (loop)

Stage Agent Routing

Each pipeline stage uses specialized agents -- not just executors. The lead selects agents based on the stage and task characteristics.

Routing rules:

1. The lead picks agents per stage, not the user. The user's N:agent-type parameter only overrides the team-exec stage worker type. All other stages use stage-appropriate specialists.
2. MCP providers complement Claude agents. Route analysis/review to Codex (ask_codex) and UI/large-context work to Gemini (ask_gemini) when available. MCP workers are one-shot and don't participate in team communication.
3. Cost mode affects model tier. In ecomode, downgrade: opus agents to sonnet, sonnet to haiku where quality permits. team-verify always uses at least sonnet.
4. Risk level escalates review. Security-sensitive or >20 file changes must include security-reviewer + code-reviewer (opus) in team-verify.

Stage Entry/Exit Criteria

• team-plan
  • Entry: Team invocation is parsed and orchestration starts.
  • Agents: explore scans codebase, planner creates task graph, optionally analyst/architect for complex tasks.
  • Exit: decomposition is complete and a runnable task graph is prepared.
• team-prd
  • Entry: scope is ambiguous or acceptance criteria are missing.
  • Agents: analyst extracts requirements, optionally product-manager/critic.
  • Exit: acceptance criteria and boundaries are explicit.
• team-exec
  • Entry: TeamCreate, TaskCreate, assignment, and worker spawn are complete.
  • Agents: workers spawned as the appropriate specialist type per subtask (see routing table).
  • Exit: execution tasks reach terminal state for the current pass.
• team-verify
  • Entry: execution pass finishes.
  • Agents: verifier + task-appropriate reviewers (see routing table).
  • Exit (pass): verification gates pass with no required follow-up.
  • Exit (fail): fix tasks are generated and control moves to team-fix.
• team-fix
  • Entry: verification found defects/regressions/incomplete criteria.
  • Agents: executor/build-fixer/debugger depending on defect type.
  • Exit: fixes are complete and flow returns to team-exec then team-verify.

Verify/Fix Loop and Stop Conditions

Continue team-exec -> team-verify -> team-fix until:

1. verification passes and no required fix tasks remain, or
2. work reaches an explicit terminal blocked/failed outcome with evidence.

team-fix is bounded by max attempts. If fix attempts exceed the configured limit, transition to terminal failed (no infinite loop).

Resume and Cancel Semantics

• Resume: restart from the last non-terminal stage using staged state + live task status.
• Cancel: /oh-my-codex:cancel requests teammate shutdown, waits for responses (best effort), marks phase cancelled with active=false, captures cancellation metadata, then deletes team resources and clears/preserves Team state per policy.
• Terminal states are complete, failed, and cancelled.

Workflow

Phase 1: Parse Input

• Extract N (agent count), validate 1-20
• Extract agent-type, validate it maps to a known OMC subagent
• Extract task description

Phase 2: Analyze & Decompose

Use explore or architect (via MCP or agent) to analyze the codebase and break the task into N subtasks:

• Each subtask should be file-scoped or module-scoped to avoid conflicts
• Subtasks must be independent or have clear dependency ordering
• Each subtask needs a concise subject and detailed description
• Identify dependencies between subtasks (e.g., "shared types must be fixed before consumers")

Phase 3: Create Team

Call TeamCreate with a slug derived from the task:

{
  "team_name": "fix-ts-errors",
  "description": "Fix all TypeScript errors across the project"
}

Response:

{
  "team_name": "fix-ts-errors",
  "team_file_path": "~/.codex/teams/fix-ts-errors/config.json",
  "lead_agent_id": "team-lead@fix-ts-errors"
}

The current session becomes the team lead (team-lead@fix-ts-errors).

Write OMC state using the state_write MCP tool for proper session-scoped persistence:

state_write(mode="team", active=true, current_phase="team-plan", state={
  "team_name": "fix-ts-errors",
  "agent_count": 3,
  "agent_types": "executor",
  "task": "fix all TypeScript errors",
  "fix_loop_count": 0,
  "max_fix_loops": 3,
  "linked_ralph": false,
  "stage_history": "team-plan"
})

Note: The MCP state_write tool transports all values as strings. Consumers must coerce agent_count, fix_loop_count, max_fix_loops to numbers and linked_ralph to boolean when reading state.

State schema fields:

Update state on every stage transition:

state_write(mode="team", current_phase="team-exec", state={
  "stage_history": "team-plan:2026-02-07T12:00:00Z,team-prd:2026-02-07T12:01:00Z,team-exec:2026-02-07T12:02:00Z"
})

Read state for resume detection:

state_read(mode="team")

If active=true and current_phase is non-terminal, resume from the last incomplete stage instead of creating a new team.

Phase 4: Create Tasks

Call TaskCreate for each subtask. Set dependencies with TaskUpdate using addBlockedBy.

// TaskCreate for subtask 1
{
  "subject": "Fix type errors in src/auth/",
  "description": "Fix all TypeScript errors in src/auth/login.ts, src/auth/session.ts, and src/auth/types.ts. Run tsc --noEmit to verify.",
  "activeForm": "Fixing auth type errors"
}

Response stores a task file (e.g. 1.json):

{
  "id": "1",
  "subject": "Fix type errors in src/auth/",
  "description": "Fix all TypeScript errors in src/auth/login.ts...",
  "activeForm": "Fixing auth type errors",
  "owner": "",
  "status": "pending",
  "blocks": [],
  "blockedBy": []
}

For tasks with dependencies, use TaskUpdate after creation:

// Task #3 depends on task #1 (shared types must be fixed first)
{
  "taskId": "3",
  "addBlockedBy": ["1"]
}

Pre-assign owners from the lead to avoid race conditions (there is no atomic claiming):

// Assign task #1 to worker-1
{
  "taskId": "1",
  "owner": "worker-1"
}

Phase 5: Spawn Teammates

Spawn N teammates using Task with team_name and name parameters. Each teammate gets the team worker preamble (see below) plus their specific assignment.

{
  "subagent_type": "oh-my-codex:executor",
  "team_name": "fix-ts-errors",
  "name": "worker-1",
  "prompt": "<worker-preamble + assigned tasks>"
}

Response:

{
  "agent_id": "worker-1@fix-ts-errors",
  "name": "worker-1",
  "team_name": "fix-ts-errors"
}

Side effects:

• Teammate added to config.json members array
• An internal task is auto-created (with metadata._internal: true) tracking the agent lifecycle
• Internal tasks appear in TaskList output -- filter them when counting real tasks

IMPORTANT: Spawn all teammates in parallel (they are background agents). Do NOT wait for one to finish before spawning the next.

Phase 6: Monitor

The lead orchestrator monitors progress through two channels:

1. Inbound messages -- Teammates send SendMessage to team-lead when they complete tasks or need help. These arrive automatically as new conversation turns (no polling needed).

2. TaskList polling -- Periodically call TaskList to check overall progress:

   #1 [completed] Fix type errors in src/auth/ (worker-1)
   #3 [in_progress] Fix type errors in src/api/ (worker-2)
   #5 [pending] Fix type errors in src/utils/ (worker-3)
   

   Format: #ID [status] subject (owner)

Coordination actions the lead can take:

• Unblock a teammate: Send a message with guidance or missing context
• Reassign work: If a teammate finishes early, use TaskUpdate to assign pending tasks to them and notify via SendMessage
• Handle failures: If a teammate reports failure, reassign the task or spawn a replacement

Task Watchdog Policy

Monitor for stuck or failed teammates:

• Max in-progress age: If a task stays in_progress for more than 5 minutes without messages, send a status check
• Suspected dead worker: No messages + stuck task for 10+ minutes → reassign task to another worker
• Reassign threshold: If a worker fails 2+ tasks, stop assigning new tasks to it

Phase 6.5: Stage Transitions (State Persistence)

On every stage transition, update OMC state:

// Entering team-exec after planning
state_write(mode="team", current_phase="team-exec", state={
  "stage_history": "team-plan:T1,team-prd:T2,team-exec:T3"
})

// Entering team-verify after execution
state_write(mode="team", current_phase="team-verify")

// Entering team-fix after verify failure
state_write(mode="team", current_phase="team-fix", state={
  "fix_loop_count": 1
})

This enables:

• Resume: If the lead crashes, state_read(mode="team") reveals the last stage and team name for recovery
• Cancel: The cancel skill reads current_phase to know what cleanup is needed
• Ralph integration: Ralph can read team state to know if the pipeline completed or failed

Phase 7: Completion

When all real tasks (non-internal) are completed or failed:

1. Verify results -- Check that all subtasks are marked completed via TaskList
2. Shutdown teammates -- Send shutdown_request to each active teammate:

   {
     "type": "shutdown_request",
     "recipient": "worker-1",
     "content": "All work complete, shutting down team"
   }
   

3. Await responses -- Each teammate responds with shutdown_response(approve: true) and terminates
4. Delete team -- Call TeamDelete to clean up:

   { "team_name": "fix-ts-errors" }
   

   Response:

   {
     "success": true,
     "message": "Cleaned up directories and worktrees for team \"fix-ts-errors\"",
     "team_name": "fix-ts-errors"
   }
   

5. Clean OMC state -- Remove .omc/state/team-state.json
6. Report summary -- Present results to the user

Agent Preamble

When spawning teammates, include this preamble in the prompt to establish the work protocol. Adapt it per teammate with their specific task assignments.

You are a TEAM WORKER in team "{team_name}". Your name is "{worker_name}".
You report to the team lead ("team-lead").

== WORK PROTOCOL ==

1. CLAIM: Call TaskList to see your assigned tasks (owner = "{worker_name}").
   Pick the first task with status "pending" that is assigned to you.
   Call TaskUpdate to set status "in_progress":
   {"taskId": "ID", "status": "in_progress", "owner": "{worker_name}"}

2. WORK: Execute the task using your tools (Read, Write, Edit, Bash).
   Do NOT spawn sub-agents. Do NOT delegate. Work directly.

3. COMPLETE: When done, mark the task completed:
   {"taskId": "ID", "status": "completed"}

4. REPORT: Notify the lead via SendMessage:
   {"type": "message", "recipient": "team-lead", "content": "Completed task #ID: <summary of what was done>", "summary": "Task #ID complete"}

5. NEXT: Check TaskList for more assigned tasks. If you have more pending tasks, go to step 1.
   If no more tasks are assigned to you, notify the lead:
   {"type": "message", "recipient": "team-lead", "content": "All assigned tasks complete. Standing by.", "summary": "All tasks done, standing by"}

6. SHUTDOWN: When you receive a shutdown_request, respond with:
   {"type": "shutdown_response", "request_id": "<from the request>", "approve": true}

== BLOCKED TASKS ==
If a task has blockedBy dependencies, skip it until those tasks are completed.
Check TaskList periodically to see if blockers have been resolved.

== ERRORS ==
If you cannot complete a task, report the failure to the lead:
{"type": "message", "recipient": "team-lead", "content": "FAILED task #ID: <reason>", "summary": "Task #ID failed"}
Do NOT mark the task as completed. Leave it in_progress so the lead can reassign.

== RULES ==
- NEVER spawn sub-agents or use the Task tool
- ALWAYS use absolute file paths
- ALWAYS report progress via SendMessage to "team-lead"
- Use SendMessage with type "message" only -- never "broadcast"

Communication Patterns

Teammate to Lead (task completion report)

{
  "type": "message",
  "recipient": "team-lead",
  "content": "Completed task #1: Fixed 3 type errors in src/auth/login.ts and 2 in src/auth/session.ts. All files pass tsc --noEmit.",
  "summary": "Task #1 complete"
}

Lead to Teammate (reassignment or guidance)

{
  "type": "message",
  "recipient": "worker-2",
  "content": "Task #3 is now unblocked. Also pick up task #5 which was originally assigned to worker-1.",
  "summary": "New task assignment"
}

Broadcast (use sparingly -- sends N separate messages)

{
  "type": "broadcast",
  "content": "STOP: shared types in src/types/index.ts have changed. Pull latest before continuing.",
  "summary": "Shared types changed"
}

Shutdown Protocol

Lead sends:

{
  "type": "shutdown_request",
  "recipient": "worker-1",
  "content": "All work complete, shutting down team"
}

Teammate receives and responds:

{
  "type": "shutdown_response",
  "request_id": "shutdown-1770428632375@worker-1",
  "approve": true
}

After approval:

• Teammate process terminates
• Teammate auto-removed from config.json members array
• Internal task for that teammate completes

IMPORTANT: The request_id is provided in the shutdown request message that the teammate receives. The teammate must extract it and pass it back. Do NOT fabricate request IDs.

MCP Workers (Hybrid Roles)

The team skill supports hybrid execution combining Claude agent teammates with external MCP workers (Codex and Gemini CLIs). Both types can make code changes -- they differ in capabilities and cost.

Execution Modes

Tasks are tagged with an execution mode during decomposition:

How MCP Workers Operate

Codex and Gemini CLIs run in full-auto mode with filesystem access. They are autonomous executors, not just analysts:

1. Lead writes task instructions to a prompt_file
2. Lead calls ask_codex or ask_gemini with working_directory set to the project root
3. The CLI reads files, makes changes, runs commands -- all within the working directory
4. Results/summary are written to output_file
5. Lead reads the output, marks the task complete, and feeds results to dependent tasks

Key difference from Claude teammates:

• MCP workers operate via CLI, not Codex's tool system
• They cannot use TaskList/TaskUpdate/SendMessage (no team awareness)
• They run as one-shot autonomous jobs, not persistent teammates
• The lead manages their lifecycle (spawn, monitor, collect results)

When to Route Where

Example: Hybrid Team with MCP Executors

/team 3:executor "refactor auth module with security review"

Task decomposition:
#1 [mcp_codex] Security review of current auth code -> output to .omc/research/auth-security.md
#2 [mcp_codex] Refactor auth/login.ts and auth/session.ts (uses #1 findings)
#3 [mcp_gemini] Redesign auth UI components (login form, session indicator)
#4 [claude_worker] Update auth tests + fix integration issues
#5 [mcp_codex] Final code review of all changes

The lead runs #1 (Codex analysis), then #2 and #3 in parallel (Codex refactors backend, Gemini redesigns frontend), then #4 (Claude teammate handles test iteration), then #5 (final Codex review).

MCP Pre-flight Analysis (Optional)

For large ambiguous tasks, run analysis before team creation:

1. Call ToolSearch("mcp") to discover deferred MCP tools (required before first use)
2. Call ask_codex (planner role) with task description + codebase context
3. Use the analysis to produce better task decomposition
4. Create team and tasks with enriched context

If ToolSearch finds no MCP tools, skip MCP pre-flight and use Claude agents instead.

This is especially useful when the task scope is unclear and benefits from external reasoning before committing to a specific decomposition.

Monitor Enhancement: Outbox Auto-Ingestion

The lead can proactively ingest outbox messages from MCP workers using the outbox reader utilities, enabling event-driven monitoring without relying solely on SendMessage delivery.

Outbox Reader Functions

readNewOutboxMessages(teamName, workerName) -- Read new outbox messages for a single worker using a byte-offset cursor. Each call advances the cursor, so subsequent calls only return messages written since the last read. Mirrors the inbox cursor pattern from readNewInboxMessages().

readAllTeamOutboxMessages(teamName) -- Read new outbox messages from ALL workers in a team. Returns an array of { workerName, messages } entries, skipping workers with no new messages. Useful for batch polling in the monitor loop.

resetOutboxCursor(teamName, workerName) -- Reset the outbox cursor for a worker back to byte 0. Useful when re-reading historical messages after a lead restart or for debugging.

Using getTeamStatus() in the Monitor Phase

The getTeamStatus(teamName, workingDirectory, heartbeatMaxAgeMs?) function provides a unified snapshot combining:

• Worker registration -- Which MCP workers are registered (from shadow registry / config.json)
• Heartbeat freshness -- Whether each worker is alive based on heartbeat age
• Task progress -- Per-worker and team-wide task counts (pending, in_progress, completed)
• Current task -- Which task each worker is actively executing
• Recent outbox messages -- New messages since the last status check

Example usage in the monitor loop:

const status = getTeamStatus('fix-ts-errors', workingDirectory);

for (const worker of status.workers) {
  if (!worker.isAlive) {
    // Worker is dead -- reassign its in-progress tasks
  }
  for (const msg of worker.recentMessages) {
    if (msg.type === 'task_complete') {
      // Mark task complete, unblock dependents
    } else if (msg.type === 'task_failed') {
      // Handle failure, possibly retry or reassign
    } else if (msg.type === 'error') {
      // Log error, check if worker needs intervention
    }
  }
}

if (status.taskSummary.pending === 0 && status.taskSummary.inProgress === 0) {
  // All work done -- proceed to shutdown
}

Event-Based Actions from Outbox Messages

This approach complements the existing SendMessage-based communication by providing a pull-based mechanism for MCP workers that cannot use Codex's team messaging tools.

Error Handling

Teammate Fails a Task

1. Teammate sends SendMessage to lead reporting the failure
2. Lead decides: retry (reassign same task to same or different worker) or skip
3. To reassign: TaskUpdate to set new owner, then SendMessage to the new owner

Teammate Gets Stuck (No Messages)

1. Lead detects via TaskList -- task stuck in in_progress for too long
2. Lead sends SendMessage to the teammate asking for status
3. If no response, consider the teammate dead
4. Reassign the task to another worker via TaskUpdate

Dependency Blocked

1. If a blocking task fails, the lead must decide whether to:
   • Retry the blocker
   • Remove the dependency (TaskUpdate with modified blockedBy)
   • Skip the blocked task entirely
2. Communicate decisions to affected teammates via SendMessage

Teammate Crashes

1. Internal task for that teammate will show unexpected status
2. Teammate disappears from config.json members
3. Lead reassigns orphaned tasks to remaining workers
4. If needed, spawn a replacement teammate with Task(team_name, name)

Team + Ralph Composition

When the user invokes /team ralph, says "team ralph", or combines both keywords, team mode wraps itself in Ralph's persistence loop. This provides:

• Team orchestration -- multi-agent staged pipeline with specialized agents per stage
• Ralph persistence -- retry on failure, architect verification before completion, iteration tracking

Activation

Team+Ralph activates when:

1. User invokes /team ralph "task" or /oh-my-codex:team ralph "task"
2. Keyword detector finds both team and ralph in the prompt
3. Hook detects MAGIC KEYWORD: RALPH alongside team context

State Linkage

Both modes write their own state files with cross-references:

// Team state (via state_write)
state_write(mode="team", active=true, current_phase="team-plan", state={
  "team_name": "build-rest-api",
  "linked_ralph": true,
  "task": "build a complete REST API"
})

// Ralph state (via state_write)
state_write(mode="ralph", active=true, iteration=1, max_iterations=10, current_phase="execution", state={
  "linked_team": true,
  "team_name": "build-rest-api"
})

Execution Flow

1. Ralph outer loop starts (iteration 1)
2. Team pipeline runs: team-plan -> team-prd -> team-exec -> team-verify
3. If team-verify passes: Ralph runs architect verification (STANDARD tier minimum)
4. If architect approves: both modes complete, run /oh-my-codex:cancel
5. If team-verify fails OR architect rejects: team enters team-fix, then loops back to team-exec -> team-verify
6. If fix loop exceeds max_fix_loops: Ralph increments iteration and retries the full pipeline
7. If Ralph exceeds max_iterations: terminal failed state

Cancellation

Cancel either mode cancels both:

• Cancel Ralph (linked): Cancel Team first (graceful shutdown), then clear Ralph state
• Cancel Team (linked): Clear Team, mark Ralph iteration cancelled, stop loop

See Cancellation section below for details.

Idempotent Recovery

If the lead crashes mid-run, the team skill should detect existing state and resume:

1. Check ~/.codex/teams/ for teams matching the task slug
2. If found, read config.json to discover active members
3. Resume monitor mode instead of creating a duplicate team
4. Call TaskList to determine current progress
5. Continue from the monitoring phase

This prevents duplicate teams and allows graceful recovery from lead failures.

Comparison: Team vs Legacy Swarm

When to use Team over Swarm: Always prefer /team for new work. It uses Codex's built-in infrastructure, requires no external dependencies, supports inter-agent communication, and has task dependency management.

Cancellation

The /oh-my-codex:cancel skill handles team cleanup:

1. Read team state via state_read(mode="team") to get team_name and linked_ralph
2. Send shutdown_request to all active teammates (from config.json members)
3. Wait for shutdown_response from each (15s timeout per member)
4. Call TeamDelete to remove team and task directories
5. Clear state via state_clear(mode="team")
6. If linked_ralph is true, also clear ralph: state_clear(mode="ralph")

Linked Mode Cancellation (Team + Ralph)

When team is linked to ralph, cancellation follows dependency order:

• Cancel triggered from Ralph context: Cancel Team first (graceful shutdown of all teammates), then clear Ralph state. This ensures workers are stopped before the persistence loop exits.
• Cancel triggered from Team context: Clear Team state, then mark Ralph as cancelled. Ralph's stop hook will detect the missing team and stop iterating.
• Force cancel (--force): Clears both team and ralph state unconditionally via state_clear.

If teammates are unresponsive, TeamDelete may fail. In that case, the cancel skill should wait briefly and retry, or inform the user to manually clean up ~/.codex/teams/{team_name}/ and ~/.codex/tasks/{team_name}/.

Configuration

Optional settings via .omc-config.json:

{
  "team": {
    "maxAgents": 20,
    "defaultAgentType": "executor",
    "monitorIntervalMs": 30000,
    "shutdownTimeoutMs": 15000
  }
}

• maxAgents - Maximum teammates (default: 20)
• defaultAgentType - Agent type when not specified (default: executor)
• monitorIntervalMs - How often to poll TaskList (default: 30s)
• shutdownTimeoutMs - How long to wait for shutdown responses (default: 15s)

Note: Team members do not have a hardcoded model default. Each teammate is a separate Codex session that inherits the user's configured model. Since teammates can spawn their own subagents, the session model acts as the orchestration layer while subagents can use any model tier.

State Cleanup

On successful completion:

1. TeamDelete handles all Codex state:
   • Removes ~/.codex/teams/{team_name}/ (config)
   • Removes ~/.codex/tasks/{team_name}/ (all task files + lock)
2. OMC state cleanup via MCP tools:

   state_clear(mode="team")
   

   If linked to Ralph:

   state_clear(mode="ralph")
   

3. Or run /oh-my-codex:cancel which handles all cleanup automatically.

IMPORTANT: Call TeamDelete only AFTER all teammates have been shut down. TeamDelete will fail if active members (besides the lead) still exist in the config.

Git Worktree Integration

MCP workers can operate in isolated git worktrees to prevent file conflicts between concurrent workers.

How It Works

1. Worktree creation: Before spawning a worker, call createWorkerWorktree(teamName, workerName, repoRoot) to create an isolated worktree at .omc/worktrees/{team}/{worker} with branch omc-team/{teamName}/{workerName}.

2. Worker isolation: Pass the worktree path as the workingDirectory in the worker's BridgeConfig. The worker operates exclusively in its own worktree.

3. Merge coordination: After a worker completes its tasks, use checkMergeConflicts() to verify the branch can be cleanly merged, then mergeWorkerBranch() to merge with --no-ff for clear history.

4. Team cleanup: On team shutdown, call cleanupTeamWorktrees(teamName, repoRoot) to remove all worktrees and their branches.

API Reference

Important Notes

• createSession() in tmux-session.ts does NOT handle worktree creation — worktree lifecycle is managed separately via git-worktree.ts
• Worktrees are NOT cleaned up on individual worker shutdown — only on team shutdown, to allow post-mortem inspection
• Branch names are sanitized via sanitizeName() to prevent injection
• All paths are validated against directory traversal

Gotchas

1. Internal tasks pollute TaskList -- When a teammate is spawned, the system auto-creates an internal task with metadata._internal: true. These appear in TaskList output. Filter them when counting real task progress. The subject of an internal task is the teammate's name.

2. No atomic claiming -- Unlike SQLite swarm, there is no transactional guarantee on TaskUpdate. Two teammates could race to claim the same task. Mitigation: The lead should pre-assign owners via TaskUpdate(taskId, owner) before spawning teammates. Teammates should only work on tasks assigned to them.

3. Task IDs are strings -- IDs are auto-incrementing strings ("1", "2", "3"), not integers. Always pass string values to taskId fields.

4. TeamDelete requires empty team -- All teammates must be shut down before calling TeamDelete. The lead (the only remaining member) is excluded from this check.

5. Messages are auto-delivered -- Teammate messages arrive to the lead as new conversation turns. No polling or inbox-checking is needed for inbound messages. However, if the lead is mid-turn (processing), messages queue and deliver when the turn ends.

6. Teammate prompt stored in config -- The full prompt text is stored in config.json members array. Do not put secrets or sensitive data in teammate prompts.

7. Members auto-removed on shutdown -- After a teammate approves shutdown and terminates, it is automatically removed from config.json. Do not re-read config expecting to find shut-down teammates.

8. shutdown_response needs request_id -- The teammate must extract the request_id from the incoming shutdown request JSON and pass it back. The format is shutdown-{timestamp}@{worker-name}. Fabricating this ID will cause the shutdown to fail silently.

9. Team name must be a valid slug -- Use lowercase letters, numbers, and hyphens. Derive from the task description (e.g., "fix TypeScript errors" becomes "fix-ts-errors").

10. Broadcast is expensive -- Each broadcast sends a separate message to every teammate. Use message (DM) by default. Only broadcast for truly team-wide critical alerts.

11. MCP workers are one-shot, not persistent -- Codex and Gemini CLIs have full filesystem access and CAN make code changes. However, they run as autonomous one-shot jobs -- they cannot use TaskList/TaskUpdate/SendMessage. The lead must manage their lifecycle: write prompt_file, call MCP, read output_file, mark task complete. They don't participate in team communication like Claude teammates do.

Converted and distributed by TomeVault — claim your Tome and manage your conversions.