Agent Skills: Team Skill

Team Skill

Spawn N coordinated agents working on a shared task list using Claude Code's implicit agent team. Claude Code 2.1.178+ removed native TeamCreate/TeamDelete; with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, each session has one implicit team and teammates are spawned directly with the Agent/Task tool using distinct name values. This skill still preserves OMC's legacy tmux/CLI worker orchestration where documented (omc team / /omc-teams).

The swarm compatibility alias was removed in #1131.

Usage

/oh-my-claudecode:team N:agent-type "task description"
/oh-my-claudecode:team "task description"
/oh-my-claudecode: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, debugger, designer, codex, gemini, antigravity). Optional; defaults to stage-aware routing. Use codex to spawn Codex CLI workers, gemini for Gemini CLI workers (enterprise/API-key tier), or antigravity for Antigravity CLI (agy) workers (Google's successor to the Gemini CLI; requires respective CLIs installed). 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:debugger "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"
# With Codex CLI workers (requires: npm install -g @openai/codex)
/team 2:codex "review architecture and suggest improvements"
# With Gemini CLI workers (requires: npm install -g @google/gemini-cli)
/team 2:gemini "redesign the UI components"
# With Antigravity CLI workers (requires: install per https://antigravity.google)
/team 2:antigravity "redesign the UI components"
# Mixed: Codex for backend analysis, Gemini/Antigravity for frontend (use /ccg instead for this)

Architecture

User: "/team 3:executor fix all TypeScript errors"
              |
              v
      [TEAM ORCHESTRATOR (Lead)]
              |
              +-- Use the session's implicit Claude Code team
              |       -> no TeamCreate call; lead remains current session
              |
              +-- Analyze & decompose task into subtasks
              |       -> explore/architect produces subtask list
              |
              +-- Create task list entries from the implementation plan
              |       -> TODO/task entries #1, #2, #3 with dependencies
              |
              +-- Update task-list entries (pre-assign owners)
              |       -> task #1 owner=worker-1, etc.
              |
              +-- Task(name="worker-1") x 3
              |       -> spawns teammates into the team
              |
              +-- Monitor loop
              |       <- teammate messages (auto-delivered by the active team surface)
              |       -> task-list/TodoWrite review for progress
              |       -> message teammates through the active team surface to unblock/coordinate
              |
              +-- Completion
                      -> request shutdown from each teammate through the active team surface
                      <- shutdown acknowledgement from teammates
                      -> clear OMC team state (no TeamDelete call)
                      -> rm .omc/state/team-state.json

Native Claude Code team model (2.1.178+):

- No per-team ~/.claude/teams/<name>/ directory is created by this skill.
- No TeamCreate/TeamDelete calls are available.
- `team_name` is accepted by native Claude Code only as ignored legacy metadata; do not rely on it for routing.
- Spawn teammates directly via Agent/Task with `name="worker-N"`.

Goal Workflow Relationship

Team is the OMC authority for parallel, staged execution. Use the deterministic conflict policies refuse, adopt_existing, and artifact_only rather than non-deterministic warning handling. If a task mentions Claude Code /goal, Ralph, UltraQA, or artifact-only Ultragoal, keep Team as the primary loop authority unless the leader explicitly hands off. Use /goal only as a documented native Claude Code handoff target or as visible evidence from the lead session; do not claim the /goal evaluator independently runs commands, reads files, or replaces team-verify / team-fix. Artifact-only Ultragoal references should be treated as durable goal ledger/checkpoint/evidence artifacts, not as worker execution by themselves.

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.

| Stage | Required Agents | Optional Agents | Selection Criteria | | --------------- | ----------------------------------- | ------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | team-plan | explore (haiku), planner (opus) | analyst (opus), architect (opus) | Use analyst for unclear requirements. Use architect for systems with complex boundaries. | | team-prd | analyst (opus) | critic (opus) | Use critic to challenge scope. | | team-exec | executor (sonnet) | executor (opus), debugger (sonnet), designer (sonnet), writer (haiku), test-engineer (sonnet) | Match agent to subtask type. Use executor (model=opus) for complex autonomous work, designer for UI, debugger for compilation issues, writer for docs, test-engineer for test creation. | | team-verify | verifier (sonnet) | test-engineer (sonnet), security-reviewer (sonnet), code-reviewer (opus) | Always run verifier. Add security-reviewer for auth/crypto changes. Add code-reviewer for >20 files or architectural changes. code-reviewer also covers style/formatting checks. | | team-fix | executor (sonnet) | debugger (sonnet), executor (opus) | Use debugger for type/build errors and regression isolation. Use executor (model=opus) for complex multi-file fixes. |

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. Specialist agents complement executor agents. Route analysis/review to architect/critic Claude agents and UI work to designer agents. Tmux CLI workers are one-shot and don't participate in team communication.
3. Cost mode affects model tier. In 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 critic.
  • Exit: acceptance criteria and boundaries are explicit.
• team-exec
  • Entry: task list 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/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).

Stage Handoff Convention

When transitioning between stages, important context — decisions made, alternatives rejected, risks identified — lives only in the lead's conversation history. If the lead's context compacts or agents restart, this knowledge is lost.

Each completing stage MUST produce a handoff document before transitioning.

The lead writes handoffs to .omc/handoffs/<stage-name>.md.

Handoff Format

## Handoff: <current-stage> → <next-stage>

- **Decided**: [key decisions made in this stage]
- **Rejected**: [alternatives considered and why they were rejected]
- **Risks**: [identified risks for the next stage]
- **Files**: [key files created or modified]
- **Remaining**: [items left for the next stage to handle]

Handoff Rules

1. Lead reads previous handoff BEFORE spawning next stage's agents. The handoff content is included in the next stage's agent spawn prompts, ensuring agents start with full context.
2. Handoffs accumulate. The verify stage can read all prior handoffs (plan → prd → exec) for full decision history.
3. On team cancellation, handoffs survive in .omc/handoffs/ for session resume. They are not deleted by native Claude Code team cleanup; no TeamDelete call exists in Claude Code 2.1.178+.
4. Handoffs are lightweight. 10-20 lines max. They capture decisions and rationale, not full specifications (those live in deliverable files like DESIGN.md).

Example

## Handoff: team-plan → team-exec

- **Decided**: Microservice architecture with 3 services (auth, api, worker). PostgreSQL for persistence. JWT for auth tokens.
- **Rejected**: Monolith (scaling concerns), MongoDB (team expertise is SQL), session cookies (API-first design).
- **Risks**: Worker service needs Redis for job queue — not yet provisioned. Auth service has no rate limiting in initial design.
- **Files**: DESIGN.md, TEST_STRATEGY.md
- **Remaining**: Database migration scripts, CI/CD pipeline config, Redis provisioning.

Resume and Cancel Semantics

• Resume: restart from the last non-terminal stage using staged state + live task status. Read .omc/handoffs/ to recover stage transition context.
• Cancel: /oh-my-claudecode: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. Handoff files in .omc/handoffs/ are preserved for potential resume.
• Terminal states are complete, failed, and cancelled.

Windows psmux tmux-compatible gate

On native Windows, do not tell users that /team requires WSL or that tmux is unavailable until the actual tmux-compatible binary has been checked. Native psmux installs a tmux-compatible command (often tmux / tmux.cmd) and is a supported Team multiplexer.

Before blocking or falling back on Windows:

1. Check tmux -V (or the platform equivalent such as where tmux followed by tmux -V).
2. Treat a successful psmux-backed tmux -V as tmux available.
3. If psmux/tmux is available, continue the normal Team flow; do not emit WSL-required guidance.
4. Only when no tmux-compatible binary is available, tell the user to install psmux for native Windows support or use WSL2 as an alternative.

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: Initialize Team State

Use the session's implicit Claude Code team. Do not call TeamCreate; Claude Code 2.1.178+ removed that tool and automatically gives the session one implicit team when CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 is enabled.

Derive a slug such as fix-ts-errors for OMC state, prompt labels, handoffs, and human-readable reporting only. Native Claude Code may accept team_name as legacy metadata, but it is ignored for routing.

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:

| Field | Type | Description | | ---------------- | ------- | --------------------------------------------------------------------------------------- | | active | boolean | Whether team mode is active | | current_phase | string | Current pipeline stage: team-plan, team-prd, team-exec, team-verify, team-fix | | team_name | string | OMC slug for state, handoffs, and reporting; ignored by native Claude Code routing | | agent_count | number | Number of worker agents | | agent_types | string | Comma-separated agent types used in team-exec | | task | string | Original task description | | fix_loop_count | number | Current fix iteration count | | max_fix_loops | number | Maximum fix iterations before failing (default: 3) | | linked_ralph | boolean | Whether team is linked to a ralph persistence loop | | stage_history | string | Comma-separated list of stage transitions with timestamps |

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

Create task list entries for each subtask using TodoWrite or the active task-list surface. Task-list tools are for tracking only; they do not create native teams.

// Task-list entry 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, update the active task-list entries 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 directly using the Agent/Task tool with distinct name values. Each teammate gets the team worker preamble (see below) plus their specific assignment. Do not call TeamCreate, and do not rely on team_name; Claude Code 2.1.178+ ignores it for native routing.

{
  "subagent_type": "oh-my-claudecode:executor",
  "name": "worker-1",
  "prompt": "<worker-preamble + assigned tasks>"
}

Response:

{
  "agent_id": "worker-1",
  "name": "worker-1"
}

Side effects:

• Teammate is spawned into the session's implicit Claude Code team
• An internal task is auto-created (with metadata._internal: true) tracking the agent lifecycle
• Internal tasks may appear in task-list 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 message team-lead when they complete tasks or need help. These arrive through the active team/conversation surface.

2. Task-list polling/review -- Periodically check TodoWrite or the active task-list surface for 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 through the active team surface
• Reassign work: If a teammate finishes early, update the task-list entry to assign pending work to them and notify through the active team surface
• 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 real tasks (non-internal) are marked completed in TodoWrite or the active task-list surface
2. Shutdown teammates -- Send shutdown_request to each active teammate through the active team surface:

   {
     "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. Clean up native team state -- Claude Code 2.1.178+ has no TeamDelete; after teammates acknowledge shutdown, clear OMC state and any local task bookkeeping.
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 OMC team "{team_name}". Your name is "{worker_name}".
You report to the team lead ("team-lead").
You are not the leader and must not perform leader orchestration actions.

== WORK PROTOCOL ==

1. CLAIM: Check TodoWrite or the active task-list surface for tasks assigned to you (owner = "{worker_name}").
   Pick the first task with status "pending" that is assigned to you.
   Mark it `in_progress` using the active task-list surface:
   {"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 through the active team/conversation surface:
   {"type": "message", "recipient": "team-lead", "content": "Completed task #ID: <summary of what was done>", "summary": "Task #ID complete"}

5. NEXT: Check TodoWrite or the active task-list surface 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 through the active team/conversation surface:
   {"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 TodoWrite or the active task-list surface 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
- NEVER run tmux pane/session orchestration commands (for example `tmux split-window`, `tmux new-session`)
- NEVER run team spawning/orchestration skills or commands (for example `$team`, `$ultrawork`, `$autopilot`, `$ralph`, `omc team ...`, `omx team ...`)
- ALWAYS use absolute file paths
- ALWAYS report progress to "team-lead" through the active team/conversation surface
- Use direct team/conversation messages with type "message" only -- never "broadcast"

Agent-Type Prompt Injection (Worker-Specific Addendum)

When composing teammate prompts, append a short addendum based on worker type:

• claude_worker: Emphasize strict TodoWrite/task-list updates, active team/conversation messages, and no orchestration commands.
• codex_worker: Emphasize CLI API lifecycle (omc team api ... --json) and explicit failure ACKs with stderr.
• gemini_worker: Emphasize bounded file ownership and milestone ACKs after each completed sub-step.
• antigravity_worker: Same expectations as gemini_worker; emphasize bounded file ownership and milestone ACKs after each completed sub-step.

This addendum must preserve the core rule: worker = executor only, never leader/orchestrator.

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 (BLOCKING)

CRITICAL: Steps must execute in exact order. Never clear OMC team state before shutdown is confirmed or timed out.

Step 1: Verify completion

Verify via TodoWrite or the active task-list surface — all real tasks (non-internal) are completed or failed.

Step 2: Request shutdown from each teammate

Lead sends:

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

Step 3: Wait for responses (BLOCKING)

• Wait up to 30s per teammate for shutdown_response
• Track which teammates confirmed vs timed out
• If a teammate doesn't respond within 30s: log warning, mark as unresponsive

Teammate receives and responds:

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

After approval, the teammate terminates or stops accepting new work. Claude Code 2.1.178+ does not expose per-team config membership or TeamDelete cleanup; track acknowledgements in OMC state/reporting instead.

Step 4: Clear OMC team state — only after ALL teammates confirmed or timed out

Claude Code 2.1.178+ has no TeamDelete. Clear OMC team state and local task bookkeeping after the blocking shutdown pass completes.

Step 5: Orphan scan for OMC tmux/CLI workers only

For legacy OMC tmux/CLI worker runs (omc team / /omc-teams), check for worker processes that survived cleanup:

node "${CLAUDE_PLUGIN_ROOT}/scripts/cleanup-orphans.mjs" --team-name fix-ts-errors

This scans for OMC worker processes matching the team name and terminates stale orphans (SIGTERM → 5s wait → SIGKILL). Supports --dry-run for inspection.

Shutdown sequence is BLOCKING: Do not clear OMC team state until all teammates have either:

• Confirmed shutdown (shutdown_response with approve: true), OR
• Timed out (30s with no response)

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.

CLI Workers (Codex and Gemini)

The team skill supports hybrid execution combining Claude agent teammates with external CLI workers (Codex CLI and Gemini CLI). Both types can make code changes -- they differ in capabilities and cost. These are standalone CLI tools, not MCP servers.

Execution Modes

Tasks are tagged with an execution mode during decomposition:

| Execution Mode | Provider | Capabilities | | --------------- | ---------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | claude_worker | Claude agent | Full Claude Code tool access (Read/Write/Edit/Bash/Task). Best for tasks needing Claude's reasoning + iterative tool use. | | codex_worker | Codex CLI (tmux pane) | Full filesystem access in working_directory. Runs autonomously via tmux pane. Best for code review, security analysis, refactoring, architecture. Requires npm install -g @openai/codex. | | gemini_worker | Gemini CLI (tmux pane) | Full filesystem access in working_directory. Runs autonomously via tmux pane. Best for UI/design work, documentation, large-context tasks. Requires npm install -g @google/gemini-cli (enterprise/API-key tier). | | antigravity_worker | Antigravity CLI (tmux pane) | Full filesystem access in working_directory. Runs autonomously via tmux pane. Same strengths as gemini_worker; Google's successor to the Gemini CLI. Install per the official instructions (agy binary). |

How CLI Workers Operate

Tmux CLI workers run in dedicated tmux panes with filesystem access. They are autonomous executors, not just analysts:

1. Lead writes task instructions to a prompt file
2. Lead spawns a tmux CLI worker with working_directory set to the project root
3. The worker reads files, makes changes, runs commands -- all within the working directory
4. Results/summary are written to an output file
5. Lead reads the output, marks the task complete, and feeds results to dependent tasks

Key difference from Claude teammates:

• CLI workers operate via tmux, not Claude Code's tool system
• They cannot use Claude Code's native task-list or team messaging surfaces
• They run as one-shot autonomous jobs, not persistent teammates
• The lead manages their lifecycle (spawn, monitor, collect results)

When to Route Where

| Task Type | Best Route | Why | | -------------------------------- | ------------------------------ | --------------------------------------------------- | | Iterative multi-step work | Claude teammate | Needs tool-mediated iteration + team communication | | Code review / security audit | CLI worker or specialist agent | Autonomous execution, good at structured analysis | | Architecture analysis / planning | architect Claude agent | Strong analytical reasoning with codebase access | | Refactoring (well-scoped) | CLI worker or executor agent | Autonomous execution, good at structured transforms | | UI/frontend implementation | designer Claude agent | Design expertise, framework idioms | | Large-scale documentation | writer Claude agent | Writing expertise + large context for consistency | | Build/test iteration loops | Claude teammate | Needs Bash tool + iterative fix cycles | | Tasks needing team coordination | Claude teammate | Needs team/conversation status updates |

Example: Hybrid Team with CLI Workers

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

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

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

Pre-flight Analysis (Optional)

For large ambiguous tasks, run analysis before team creation:

1. Spawn Task(subagent_type="oh-my-claudecode:planner", ...) with task description + codebase context
2. Use the analysis to produce better task decomposition
3. Create team and tasks with enriched context

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 CLI workers using the outbox reader utilities, enabling event-driven monitoring alongside native team/conversation 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

| Message Type | Action | | --------------- | ------------------------------------------------------------------------------------------- | | task_complete | Mark task completed, check if blocked tasks are now unblocked, notify dependent workers | | task_failed | Increment failure sidecar, decide retry vs reassign vs skip | | idle | Worker has no assigned tasks -- assign pending work or begin shutdown | | error | Log the error, check consecutiveErrors in heartbeat for quarantine threshold | | shutdown_ack | Worker acknowledged shutdown -- safe to remove from team | | heartbeat | Update liveness tracking (redundant with heartbeat files but useful for latency monitoring) |

This approach complements native team/conversation messaging by providing a pull-based mechanism for MCP workers that cannot use Claude Code's team messaging tools.

Error Handling

Teammate Fails a Task

1. Teammate reports the failure to the lead through the active team/conversation surface
2. Lead decides: retry (reassign same task to same or different worker) or skip
3. To reassign: update the active task-list entry with the new owner, then message the new owner through the active team surface

Teammate Gets Stuck (No Messages)

1. Lead detects via TodoWrite or the active task-list surface -- task stuck in in_progress for too long
2. Lead messages the teammate asking for status through the active team surface
3. If no response, consider the teammate dead
4. Reassign the task to another worker through the active task-list surface

Dependency Blocked

1. If a blocking task fails, the lead must decide whether to:
   • Retry the blocker
   • Remove the dependency by updating the active task-list entry's blockedBy metadata
   • Skip the blocked task entirely
2. Communicate decisions to affected teammates through the active team surface

Teammate Crashes

1. Internal tracking for that teammate will show unexpected status
2. Lead reassigns orphaned tasks to remaining workers
3. If needed, spawn a replacement teammate with Task(name="worker-N", subagent_type="...")

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-claudecode: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-claudecode: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 OMC state and resume:

1. Read state_read(mode="team") for the active OMC team slug, phase, and worker labels
2. Use OMC handoffs and task-list/TodoWrite state to determine current progress
3. Resume monitor mode instead of spawning duplicate teammates
4. Continue from the last recorded stage

This prevents duplicate worker spawns and allows graceful recovery from lead failures.

Comparison: Team vs Legacy Swarm

| Aspect | Team (Native Claude Code 2.1.178+) | Swarm (Legacy SQLite) | | ----------------------- | ---------------------------------------------------------------- | -------------------------------------- | | Storage | OMC state/handoffs plus Claude Code's current task-list surface | SQLite in .omc/state/swarm.db | | Dependencies | better-sqlite3 not needed | Requires better-sqlite3 npm package | | Task claiming | Lead pre-assigns named workers through task-list/TodoWrite state | SQLite IMMEDIATE transaction -- atomic | | Race conditions | Possible if two agents claim same task (mitigate by pre-assigning) | None (SQLite transactions) | | Communication | Native implicit-team messages / conversation turns | None (fire-and-forget agents) | | Task dependencies | Lead-managed dependencies in task-list/TodoWrite state | Not supported | | Heartbeat | Lead detects via missing messages/status | Manual heartbeat table + polling | | Shutdown | Graceful request/response protocol plus OMC state clear | Signal-based termination | | Agent lifecycle | Tracked by named Agent/Task spawns and OMC state | Manual tracking via heartbeat table | | Progress visibility | Task list/TodoWrite state with named worker ownership | SQL queries on tasks table | | Conflict prevention | Owner labels (lead-assigned) | Lease-based claiming with timeout | | Crash recovery | Lead detects via missing messages, reassigns | Auto-release after 5-min lease timeout | | State cleanup | Clear OMC team state after teammate shutdown | Manual rm of SQLite database |

When to use Team over Swarm: Always prefer /team for new native Claude Code work. It uses Claude Code's implicit agent team, requires no external dependencies, supports inter-agent coordination, and has task dependency management.

Cancellation

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

1. Read team state via state_read(mode="team") to get team_name and linked_ralph
2. Request shutdown from all active named teammates through the active team surface
3. Wait for shutdown_response from each (15s timeout per member)
4. Clear state via state_clear(mode="team")
5. 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, record the timeout, avoid spawning more work, and clear OMC state only after the shutdown wait completes or the user force-cancels.

Runtime V2 (Event-Driven)

When OMC_RUNTIME_V2=1 is set, the team runtime uses an event-driven architecture instead of the legacy done.json polling watchdog:

• No done.json: Task completion is detected via CLI API lifecycle transitions (claim-task, transition-task-status)
• Snapshot-based monitoring: Each poll cycle takes a point-in-time snapshot of tasks and workers, computes deltas, and emits events
• Event log: All team events are appended to .omc/state/team/{teamName}/events.jsonl
• Worker status files: Workers write status to .omc/state/team/{teamName}/workers/{name}/status.json
• Preserved: Sentinel gate (blocks premature completion), circuit breaker (dead worker detection), failure sidecars

The v2 runtime is feature-flagged and can be enabled per-session. The legacy v1 runtime remains the default.

Dynamic Scaling

When OMC_TEAM_SCALING_ENABLED=1 is set, the team supports mid-session scaling:

• scale_up: Add workers to a running team (respects max_workers limit)
• scale_down: Remove idle workers with graceful drain (workers finish current task before removal)
• File-based scaling lock prevents concurrent scale operations
• Monotonic worker index counter ensures unique worker names across scale events

Configuration

Optional settings live in .claude/omc.jsonc (project) or ~/.config/claude-omc/config.jsonc (user). Project values override user values; OMC_TEAM_ROLE_OVERRIDES (env JSON) supersedes both.

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

• ops.maxAgents - Maximum teammates (default: 20)
• ops.defaultAgentType - CLI provider when a /team invocation does not specify one (claude | codex | gemini | antigravity | grok | cursor, default: claude)
• ops.monitorIntervalMs - How often to review TodoWrite or the active task-list surface (default: 30s)
• ops.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 Claude Code 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.

Per-Role Provider & Model Routing

Scope: Applies to /team only. Task-based delegation uses delegationRouting (see separate docs). The two systems coexist by design.

Declare which provider (claude, codex, gemini, antigravity, grok, cursor) and which model tier should back each canonical role. Routing is resolved once at team creation and persisted in TeamConfig.resolved_routing — spawn, scale-up, and restart all read from the snapshot, so a role's worker CLI and model are stable for the lifetime of the team.

Example — user target mapping

// .claude/omc.jsonc
{
  "team": {
    "roleRouting": {
      "orchestrator": { "model": "inherit" },
      "planner": { "provider": "claude", "model": "HIGH" },
      "analyst": { "provider": "claude", "model": "HIGH" },
      "executor": { "provider": "claude", "model": "MEDIUM" },
      "debugger": { "provider": "cursor" },
      "critic": { "provider": "codex" },
      "code-reviewer": { "provider": "gemini" },
      "test-engineer": { "provider": "gemini", "model": "MEDIUM" },
    },
  },
}

| Role | Provider | Model | | --------------- | --------------- | ------------------------- | | orchestrator | claude (pinned) | inherits invoking session | | planner | claude | HIGH (opus) | | analyst | claude | HIGH (opus) | | executor | claude | MEDIUM (sonnet) | | debugger | cursor | cursor-agent default | | critic | codex | codex default | | code-reviewer | gemini | gemini default | | test-engineer | antigravity | antigravity default |

Canonical roles

orchestrator, planner, analyst, architect, executor, debugger, critic, code-reviewer, security-reviewer, test-engineer, designer, writer, code-simplifier, explore, document-specialist.

User-friendly aliases normalize via normalizeDelegationRole() — e.g. reviewer → code-reviewer, quality-reviewer → code-reviewer, harsh-critic → critic, build-fixer → debugger. Accepted alias keys are honored during resolved snapshot creation and later stage routing, not just validation. Unknown roles fail validation at parse time.

Spec fields (TeamRoleAssignmentSpec)

• provider — "claude" | "codex" | "gemini" | "antigravity" | "grok" | "cursor". Omitted → defaults to claude.
• model — tier name ("HIGH" | "MEDIUM" | "LOW") or an explicit model ID. Tiers resolve through routing.tierModels.
• agent — optional Claude agent name (e.g. "critic", "executor"). Only honored when the resolved provider is claude.

orchestrator is pinned to claude; only model is user-configurable. Any other key on orchestrator is rejected by the validator.

cursor launches cursor-agent as an interactive executor/refactor worker. Do not route reviewer/verdict roles (critic, code-reviewer, security-reviewer, test-engineer) to Cursor unless its CLI gains a compatible verdict-output mode; the runtime intentionally skips the structured verdict contract for Cursor panes.

Env override

OMC_TEAM_ROLE_OVERRIDES='{"critic":{"provider":"codex"},"code-reviewer":{"provider":"gemini"}}'

Precedence: OMC_TEAM_ROLE_OVERRIDES > .claude/omc.jsonc (project) > ~/.config/claude-omc/config.jsonc (user) > built-in defaults. Invalid JSON logs a warning and is ignored — env overrides are best-effort and never abort the run.

Fallback when a CLI is missing

If the CLI for a configured provider is absent from PATH at spawn time, buildLaunchArgs() throws, the team lead emits a visible team/conversation warning, and the runtime falls back to a deterministic Claude assignment pre-computed by buildResolvedRoutingSnapshot (same tier + same agent, provider: "claude"). Fallback is loud by design — silent fallback is a test failure. Probe provider availability with omc doctor --team-routing.

Stickiness — resolved once, reused everywhere

Resolved routing is immutable per team. Editing config mid-team-lifetime does not affect running teams; a new /team invocation picks up the new mapping. This guarantees that spawn, scale-up, and worker-restart all see identical routing, including across worktree detaches (the snapshot travels with TeamConfig).

Zero-config behavior

An empty team.roleRouting preserves pre-patch behavior: every worker is Claude, model tiers follow routing.tierModels, and /team 3:executor ... still spawns three Claude Sonnet executors.

State Cleanup

On successful completion:

1. Native Claude Code 2.1.178+ has no per-team TeamDelete cleanup. After shutdown is confirmed or timed out, clear OMC state via MCP tools:

   state_clear(mode="team")
   

   If linked to Ralph:

   state_clear(mode="ralph")
   

2. For legacy OMC tmux/CLI workers, run the documented omc team shutdown / cleanup path.
3. Or run /oh-my-claudecode:cancel which handles OMC state cleanup automatically.

IMPORTANT: Clear OMC team state only AFTER all teammates have been shut down or timed out.

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

| Function | Description | | ------------------------------------------------------------------- | ------------------------------ | | createWorkerWorktree(teamName, workerName, repoRoot, baseBranch?) | Create isolated worktree | | removeWorkerWorktree(teamName, workerName, repoRoot) | Remove worktree and branch | | listTeamWorktrees(teamName, repoRoot) | List all team worktrees | | cleanupTeamWorktrees(teamName, repoRoot) | Remove all team worktrees | | checkMergeConflicts(workerBranch, baseBranch, repoRoot) | Non-destructive conflict check | | mergeWorkerBranch(workerBranch, baseBranch, repoRoot) | Merge worker branch (--no-ff) | | mergeAllWorkerBranches(teamName, repoRoot, baseBranch?) | Merge all completed workers |

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/lifecycle task entries may pollute task-list output -- If Claude Code reports internal lifecycle entries for spawned teammates, filter them when counting real task progress. The subject of an internal task is often the teammate's name.

2. No atomic claiming -- Unlike SQLite swarm, native task-list/TodoWrite state does not provide transactional claiming. Two teammates could race to claim the same task. Mitigation: The lead should pre-assign owners before spawning teammates. Teammates should only work on tasks assigned to them.

3. Task IDs are strings when exposed by task-list tools -- IDs may be auto-incrementing strings ("1", "2", "3"), not integers. Always pass string values to taskId fields when using task-list tools.

4. No TeamDelete cleanup -- Claude Code 2.1.178+ removed TeamDelete; use shutdown messages plus OMC state cleanup.

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. Do not put secrets in teammate prompts -- Prompts can be retained in logs, state, or conversation history. Keep credentials and sensitive data out of teammate prompts.

7. Shutdown acknowledgements are state/reporting events -- After a teammate approves shutdown and terminates, track that acknowledgement in OMC state/reporting. Do not expect a Claude Code team membership config to update.

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. CLI workers are one-shot, not persistent -- Tmux CLI workers have full filesystem access and CAN make code changes. However, they run as autonomous one-shot jobs -- they cannot use Claude Code's native task-list or team messaging surfaces. The lead must manage their lifecycle: write prompt_file, spawn CLI worker, read output_file, mark task complete. They don't participate in team communication like Claude teammates do.

Parallel session caveats

• Multi-repo workspace anchor: drop a .omc-workspace marker at the parent directory so multiple sessions across sub-repos share one .omc/. Resolution order: OMC_STATE_DIR > .omc-workspace > git > cwd. See docs/REFERENCE.md.
• Session id source: OMC_SESSION_ID env var wins in CLI contexts; hook payload data.session_id wins in hook contexts.
• Plan id (when applicable): Team state is session-scoped. Team handoffs at .omc/handoffs/ are shared by design (see Wave G in the workspace plan).
• Parallel verdict: supported (session-scoped + shared handoffs by design)