Multi Agent Teammode

Teammode

Run a named team of cooperating Codex workers under one leader, with durable state on disk. This is a Codex-only workflow. It never depends on an external terminal runner - it coordinates through Codex's own collaboration tools plus a bundled state script, on ONE of two transports chosen up front: native MultiAgentV2 agents, or Codex App threads as the fallback.

When to use a team (and when to use plain subagents instead)

Use a TEAM when EITHER holds:

• the work does NOT split into perfectly isolated pieces, but doing it in parallel is clearly more convenient - members will need to see and react to each other's findings; or
• one task still needs exploration, yet its GOAL is already clear - parallel investigation under a fixed objective.

Use plain fire-and-forget subagents ($ulw / one-off spawn_agent workers) - NOT a team - when EITHER holds:

• the work IS perfectly isolated, so there is no coordination cost worth paying; or
• the GOAL is still ambiguous, where one mind should resolve direction before any fan-out.

A team buys cross-member coordination at a real overhead cost; only spend it when coordination is the thing you actually need.

Pick the transport FIRST - then tell the user

Before creating any team state, decide which transport this session can run. Inspect your active tool list and select:

1. MultiAgentV2 (preferred) - select when the flat V2 collaboration tools are ALL active: flat spawn_agent whose schema requires task_name, plus send_message, followup_task, wait_agent, list_agents, and interrupt_agent. Members are durable native agents addressed by task name / agent path (/root/<task_name>). The namespaced multi_agent_v1.* surface never qualifies as a team transport.
2. Codex App threads (fallback) - select when flat V2 is not available but the codex_app.* thread tools are (create_thread, read_thread, send_message_to_thread, set_thread_title, set_thread_archived).
3. Neither set visible - if a tool_search tool is active, search for the missing sets (e.g. spawn_agent, codex_app) before concluding: some environments defer tools behind tool search. A hit is only a lead: revalidate that the visible result is the COMPLETE, mutually compatible transport set from case 1 or 2 before selecting it. Do not combine partial hits from different transports.
4. Neither set exists - teammode cannot run here. Do NOT run init or fake a team with partial tooling. If another visible plain-subagent mechanism can independently spawn, communicate with, and observe plain workers, announce that exact mechanism and use it for non-overlapping scopes. Otherwise continue serially and report the capability limitation; never promise or imply plain subagents that this session cannot create.

Then, BEFORE running init (or instead of it in case 4), tell the user in one line what this environment provides and which route you picked:

• Teammode transport: MultiAgentV2 (flat spawn_agent with task_name).
• Teammode transport: Codex App threads (flat V2 tools not present in this session).
• Teammode unavailable: neither MultiAgentV2 nor codex_app tools exist in this session - using <visible plain-subagent mechanism> for independent scopes.
• Teammode unavailable: neither MultiAgentV2 nor codex_app tools exist in this session, and no compatible plain-subagent mechanism is available - continuing serially.

Pass the choice to init as --transport multi_agent_v2 or --transport codex_app. The transport is recorded in team.json and is IMMUTABLE for the team's lifetime: a V2 spawn failure is a V2 blocker to report, never permission to mix Codex App threads into the same team. Never probe by trial-calling tools; read your tool list, and search it with tool_search only when a needed set is not visible.

You are the leader - orchestrate, do not implement

The main session is ALWAYS the team leader; you orchestrate directly and never spin up a separate leader worker. Your job is orchestration, NOT writing product code: split the work and assign each slice, hold live situational awareness of every member, verify and QA what they deliver, relay findings between members, instruct and unblock, and synthesize the result. DELEGATE every code edit to a member - if you catch yourself editing product files while the team runs, that work was a member's slice you should have handed off. You own direction, verification, and integration (the merge), not the keystrokes.

Compose by part, ownership, or perspective - not by job title

A team is ALWAYS two or more members - never a single-member team. One worker on an isolated job is a plain subagent, not a team; if you end up with a single member, either split off a second distinct slice or drop the team and use a subagent.

Compose the team from what you actually KNOW about the work. Ground the split in real knowledge of the problem, then divide it into clear, non-overlapping responsibilities - one per aspect of the work - and give each member exactly one. No two members may own the same thing. Define each member by a concrete slice: a specific part of the codebase, an ownership area, or a distinct perspective/lens. Assigning a vague role ("backend dev", "release analyst", "the tester") is an anti-pattern - it gives the member no real boundary and invites overlap. Each member's focus names what they own concretely; the lens is one of area, ownership, or perspective. Give each member a short, distinct --name too - its role or what it watches (e.g. app-server-lifecycle, mailbox-delivery) - it labels the member everywhere; never reuse one name for two members. On MultiAgentV2 teams also give each member a unique --task-name in lowercase_digits_underscores form - it becomes the member's permanent agent path /root/<task_name>.

Run the script - never hand-write team state

A bundled, dependency-free Node script owns all team state so you never author team.json or the member manual by hand. Run it with node (or bun); it works on macOS, Linux, and Windows. Replace <skill-root> with this skill's own directory.

node "<skill-root>/scripts/team.mjs" init        --name "<team>" --session-name "<session>" --transport multi_agent_v2|codex_app [--session <leader_thread_id>] [--worktree] [--base-branch dev]
node "<skill-root>/scripts/team.mjs" add-member  --team <session_id> --id A --name "<short role>" [--task-name <v2_task_name>] --focus "<part/ownership/perspective>" --lens area|ownership|perspective --deliverable "<...>" [--branch <branch>]
node "<skill-root>/scripts/team.mjs" bind-agent   --team <session_id> --id A --agent-path /root/<task_name> [--cwd <path>]   # multi_agent_v2 teams
node "<skill-root>/scripts/team.mjs" bind-thread  --team <session_id> --id A --thread <thread_id> [--cwd <path>]             # codex_app teams
node "<skill-root>/scripts/team.mjs" member-prompt --team <session_id> --id A
node "<skill-root>/scripts/team.mjs" set-status   --team <session_id> --id A --status reported|blocked|active|archived [--note "<...>"]
node "<skill-root>/scripts/team.mjs" worktree-add    --team <session_id> --id A [--base-branch <branch>]
node "<skill-root>/scripts/team.mjs" worktree-remove --team <session_id> --id A [--force]
node "<skill-root>/scripts/team.mjs" integrate       --team <session_id> [--id A]
node "<skill-root>/scripts/team.mjs" archive      --team <session_id> [--id A] [--note "<...>"]
node "<skill-root>/scripts/team.mjs" delete       --team <session_id> [--force]
node "<skill-root>/scripts/team.mjs" status       --team <session_id>

init creates .omo/teams/{session_id}/ containing team.json (the single durable state file: team id, transport, the main-session leader, the member roster, status, worktree config, and a lifecycle log), guide.md (the auto-generated member field manual), and artifacts/ (a shared exchange space). On codex_app teams {session_id} is the leader's Codex session id when you can pass it via --session; otherwise the script generates a stable handle. Re-running init is a safe no-op. Every mutating subcommand rewrites guide.md, so the manual always matches the current team.

Mutating subcommands take a per-team state lock before reading and rewriting team.json. It is safe to run independent add-member, bind-agent, bind-thread, set-status, archive, delete, guide, and worktree mutation commands concurrently against the same team: they serialize and each command reads the latest committed state before writing. If a command reports that team state is locked, do not treat the intended mutation as complete; retry after the named command finishes, or inspect .omo/teams/{session_id}/.team.lock/owner.json if the previous command crashed. bind-agent refuses codex_app teams and bind-thread refuses multi_agent_v2 teams, and a refused command never changes team.json.

Create the team and its members

init the team, then add-member once per member. What happens next depends on the transport.

MultiAgentV2 teams:

1. If a member needs an isolated worktree, run worktree-add BEFORE spawning it - flat spawn_agent has no cwd argument, so the path must ride in the bootstrap message.
2. Spawn each member with flat spawn_agent using only the V2 schema fields: task_name is that member's --task-name, message is the bootstrap printed by add-member / member-prompt, and fork_turns is "none" (members read guide.md for context; full parent history is not their context model). Put any role, priority, or task-specific routing instruction in message; V2 does not accept agent_type, model, reasoning_effort, or service_tier, so members inherit the session model.
3. bind-agent --agent-path with the canonical task name the spawn returned (normally /root/<task_name>); binding confirms the runtime identity matches the roster and records the member's cwd. Members are durable: they persist as subagent threads, survive idling, and are re-tasked with followup_task - never respawned under a second name.
4. Members appear in list_agents with their task paths; inspect status there instead of deep links (V2 exposes no thread title or codex:// link surface to you).

Codex App teams:

1. Create a durable thread per member with codex_app.create_thread - ALWAYS this tool for every member - titled [team name] <member name>, using THAT member's own name, so no two threads share a title. add-member prints the exact title to use. If the tool accepts a working directory / cwd argument, set it to that member's worktree; otherwise the member's manual tells it to cd there first. Use codex_app.set_thread_title if the title did not land at creation. If Codex returns only pendingWorktreeId, the worktree-backed thread is not ready yet: do not bind-thread and do not send the member bootstrap. Wait until Codex surfaces a real threadId, then set the title, bind that real id with the cwd, and only then send the bootstrap.
2. bind-thread to record each thread id (and --cwd), then send that member's bootstrap trigger as the thread's first message. The trigger is short on purpose: it tells the new thread to READ its guide.md and team.json rather than carrying the whole protocol inline.
3. Whenever you report, audit, reopen, or hand off a member thread, include the app deep link codex://threads/<thread_id> next to the raw id - worktree-backed threads are easy to lose in the sidebar without it. On a codex_app team, a spawned in-process agent is never a member substitute: it cannot carry the team title or be inspected, titled, archived, or re-opened with the codex_app.* tools this team runs on.

On either transport, a member only counts once it is bound (bind-agent / bind-thread). If the selected transport's tools stop working mid-run, STOP and say so (see Stop rules); do not quietly switch transports.

Communication

Members push to you and to one another; you never poll them as a routine. The address book is team.json, and the generated manual binds members to the hard rules, so you mainly keep the channel open: expect frequent small inbound updates from each member - findings, WORKING:/BLOCKED: markers, peer digests - rather than one final dump, and act on them as they arrive.

• MultiAgentV2: members reach you with send_message to /root and reach peers by their members[].agentPath. You reach members the same way; use followup_task when you hand an idle member NEW work (it wakes the member), send_message for context that should not interrupt, and wait_agent only when you are genuinely blocked on their next update - a wait_agent timeout only means no new mailbox update arrived, never that a member failed. Your own session IS /root - members can always reach you; leave --session unset.
• Codex App: members push with codex_app.send_message_to_thread; you inspect state with codex_app.read_thread. So members can actually reach you, run init with --session <your own thread id> - that makes leader.sessionId in team.json a real, messageable thread; without it members cannot report to you and you are stuck polling.

All member-to-member and member-to-leader traffic is in English; when the END user addresses a member, that member replies in the user's own language. Members hand off files and memos through the team artifacts/ directory and reference them by path.

Let members work - do not rush them

Members heartbeat every few tool calls and message you on every finding, blocker, and finished slice (their manual binds them to this). So a member that is quiet between heartbeats is working, not stalled - a stretch of silence is the normal sound of focused work, not a problem to chase. Re-reading a calm member's state, or sending "any update?" / "are you done?" / "hurry up" pings, interrupts that member and slows the whole team. Trust the heartbeat and let them cook.

Message a member only when one of these is true:

• you have new information, context, or a correction it needs to do its slice right;
• you are reassigning, narrowing, or unblocking its scope;
• a peer's result changes what it should do; or
• it has gone fully silent well past its heartbeat cadence AND that stall is blocking the team - then send one specific question, not a barrage.

Otherwise stay calm and keep the channel open: read inbound updates as they arrive and act on them. A long-running member is alive; a heartbeat you have not received yet is not a failure. Fallback only when a member is completed without its deliverable, explicitly BLOCKED:, or no longer running - then unblock, reassign, or re-task that slice instead of waiting on it. Wait for every required member's final report before you declare the team done - rushing toward "done" while members are still mid-slice just produces half-built work you will have to redo.

Worktrees - isolate members who would touch the same files

The moment two members' slices would edit the same files, give each its own git worktree so they cannot clobber each other. Decide this whenever you see the collision - at team creation OR mid-run, not only up front. For each colliding member run worktree-add --team <id> --id <member>: it creates the worktree off the base branch on a derived branch, flips the team into worktree mode, records it in team.json, and prints the cd path to hand that member. The member works and commits only inside its own worktree. To land the work, integrate --team <id> merges every member branch into your current branch with a merge commit (never a squash or rebase); resolve any conflict it reports, then worktree-remove each worktree at cleanup.

Delivering the path differs by transport: on MultiAgentV2, create the worktree BEFORE spawning so the bootstrap carries it, or send it to an already-running member as a followup_task; on Codex App, send a follow-up message that includes both the worktree path and the member's codex://threads/<thread_id> link. Either way, run bind-agent/bind-thread with --cwd <worktree> (or --worktree-path) so team.json, guide.md, status, and member-prompt all point at the same worktree-backed member.

When the member starts inside a worktree, it must verify the assigned cwd exists and contains the repository checkout before editing. If the directory is missing, empty, or does not look like a git worktree/repository yet, the member reports BLOCKED: worktree not ready to the leader and waits instead of editing a parent checkout or an empty directory.

Run a ulw-plan in parallel

When a decision-complete plan already exists at .omo/plans/<slug>.md (from ulw-plan), execute its parallel waves as a team instead of one todo at a time. Map it directly:

• one wave's independent todos -> one member each; the todo's scope/files become that member's focus, and its acceptance criteria + QA become the member's deliverable.
• the plan's dependency matrix sets the shape: todos with no unmet dependency inside a wave run as concurrent members; a todo that depends on another waits, so launch the next wave only after the blocking members report.
• todos in the same wave that touch overlapping files -> give those members worktrees (see above).

Keep the plan file as the shared spec: point each member at its todo by path, and verify the member's result against that todo's acceptance criteria before you integrate.

Archive, delete, and cleanup

DISBAND the team the moment it is no longer needed. A team exists only to do its work; once that work is done, or the user no longer wants it, do not leave it lying around - archive every member, then delete the team state only after archival evidence is clean or preserved. A finished team that is never disbanded is a leak.

• archive closes the team: notify each active member, copy anything useful into artifacts/, then close each member on its transport. On MultiAgentV2, interrupt_agent any member still mid-turn and record in the note that V2 exposes no runtime archive operation - the durable team.json state IS the archive; never claim a V2 agent itself was archived. On Codex App, try codex_app.set_thread_archived per member thread; treat failures such as "Ambiguous Codex thread id" or an id that is ambiguous across hosts as an app-thread archival blocker, not as a team-state blocker: record the failure in the team log, tell the user which member thread was not proven archived, and continue the team-state archive with archive --note "<blocker>". Never pretend a member thread was archived. Do not delete the team state after an app-thread archival blocker unless the evidence has been copied elsewhere or the user explicitly accepts that evidence loss.
• delete removes .omo/teams/{session_id} and refuses while the team is unarchived or any member is still active unless --force.
• When the work wraps up, land it the way the user asked: integrate --team <id> for a direct merge commit, or push each member branch and open a PR. Then worktree-remove each worktree, archive, and delete. Cleanup is real work; respect the user's instruction on how to land it.

Stop rules

• Stop and ask before deleting an unarchived team while any member is still active.
• Member communication stays English unless the user explicitly requests otherwise; user-facing replies follow the user's language.
• Stop if the selected transport's tools (V2 spawn/message/wait/list/interrupt, or Codex App create/read/send/title/archive) stop working mid-run; say so instead of faking it or silently switching transports. Pre-init absence is not a stop - it routes to the plain-subagent split.

Execution Workflow

1. Initialize isolated execution context.
2. Run targeted capabilities.
3. Produce verifiable structured evidence.