Skip to content

Reference: `team-compose.yaml`

The compose tree has two layers:

One global file (team-compose.yaml) — broker, supervisor, budget, HITL policy, interfaces, and the list of project files.
One per-project file (projects/<id>.yaml) — channels, managers, workers.

Global

version: "2.0.0"

broker:
  type: sqlite                        # sqlite (default); redis-streams is planned
  path: state/mailbox.db

supervisor:
  type: tmux                          # tmux (default) | systemd | launchd
  tmux_prefix: a-                     # tmux session name prefix

budget:
  daily_usd_limit: 25.0
  warn_threshold_pct: 75
  message_ttl_hours: 24
  per_project_usd_limit:
    newsroom: 15.0

hitl:
  globally_sensitive_actions:
    - publish
    - release
    - deploy
    - payment
    - external_email
    - external_api_post
    - merge_to_main
    - dns_change
  auto_approve_windows:
    - action: publish
      project: newsroom
      scope: "morning-brief-*"
      until: "2026-05-01T09:00:00Z"

projects:
  - file: projects/newsroom.yaml
  - file: projects/blog-site.yaml

Telegram bots live on the manager definition itself — see the per-project example below — and are configured via teamctl bot setup, which writes both the env vars and the telegram: block. The top-level interfaces: array is no longer needed for Telegram.

The example uses the implemented tmux backend. Schema fields can also describe planned backends and policy metadata; the field table below distinguishes those from enforced behavior.

Per-project

version: 2

project:
  id: newsroom
  name: Newsroom
  cwd: .

channels:
  - name: editorial
    members: [head_editor, fact_checker, news_writer]
  - name: all
    members: "*"

managers:
  head_editor:
    runtime: claude-code
    model: claude-opus-4-8
    role_prompt: roles/head_editor.md
    permission_mode: auto
    autonomy: low_risk_only
    can_dm: [fact_checker, news_writer, seo]
    can_broadcast: [editorial, all]
    # Per-manager 1:1 Telegram bot. Run `teamctl bot setup` to populate
    # both the env vars and this block. After setup, `teamctl up`
    # spawns one team-bot per manager and DMing the bot reaches the
    # manager directly — no `/dm role text` needed.
    interfaces:
      telegram:
        bot_token_env: TEAMCTL_TG_HEAD_EDITOR_TOKEN
        chat_ids_env: TEAMCTL_TG_HEAD_EDITOR_CHATS

workers:
  fact_checker:
    runtime: gemini
    model: gemini-3.0-pro
    role_prompt: roles/fact_checker.md
    reports_to: head_editor
    can_dm: [head_editor, news_writer]
    can_broadcast: [editorial]

  news_writer:
    runtime: claude-code
    model: claude-sonnet-4-6
    role_prompt: roles/news_writer.md
    reports_to: head_editor
    can_dm: [head_editor, fact_checker]
    can_broadcast: [editorial]

Field reference

Global

Field	Type	Default	Notes
`version`	string	—	Schema version, the semver string `"2.0.0"`. A legacy integer `2` is also accepted and self-heals (rewritten to `"2.0.0"` on the next save).
`broker.type`	string	`sqlite`	Only `sqlite` is shipping.
`broker.path`	string	`state/mailbox.db`	Resolved relative to the compose root.
`supervisor.type`	string	`tmux`	Only `tmux` is implemented. `systemd` and `launchd` are planned.
`supervisor.tmux_prefix`	string	`a-`	Tmux session name = `<prefix><project>-<agent>`.
`budget.daily_usd_limit`	float	—	Intended overall limit; not an enforced provider-spend ceiling today.
`budget.per_project_usd_limit`	map	`{}`	Intended per-project limits; no runtime cost parser or spend enforcement today.
`budget.message_ttl_hours`	int	24	`teamctl gc` horizon.
`hitl.globally_sensitive_actions`	list	(see default)	Policy metadata for role instructions; does not intercept arbitrary runtime tools.
`hitl.auto_approve_windows`	list	`[]`	Declared policy windows; automatic approval is not implemented.
`interfaces`	list	`[]`	Reserved for non-Telegram adapters (discord, imessage, cli, webhook, email). Telegram now lives on the manager.
`projects`	list	`[]`	Each entry: `{ file: <path> }`.

Per-project

Field	Type	Default	Notes
`version`	int	—	Project schema version (the integer `2`).
`project.id`	string	—	Unique id; used in tmux session names, mailbox scoping.
`project.name`	string	—	Human label.
`project.cwd`	path	—	Working directory for runtimes. Relative paths resolve against the compose root.
`channels[].name`	string	—	Channel name (project-scoped).
`channels[].members`	list or `"*"`	—	Agent short-names or `"*"` for every agent in this project.
`managers` / `workers`	map	—	Keyed by agent short-name.

Agent

Field	Type	Default	Notes
`runtime`	string	`claude-code`	Must match a `runtimes/<name>.yaml`.
`model`	string	runtime default	Runtime-specific model id.
`role_prompt`	path or list of paths	—	System prompt source. A single string keeps the current behavior — the file is passed straight to the runtime. A list concatenates the files in declared order (separated by an em-dash) into `state/role_prompts/<project>-<agent>.md` and points the runtime at that file. Re-materialized on every render, so editing any source file flows into the agent’s prompt at the next `up` / `reload`.
`permission_mode`	string	`auto` (headless)	Claude Code permission mode. Headless agents default to `auto`: a classifier lets routine work run without prompts and blocks risky actions outright, so an unattended pane never freezes on a permission dialog. `attended` is teamctl-specific — it marks a human-at-keyboard agent, so `teamctl` passes no `--permission-mode` and skips the `PreToolUse` deny hook that blocks synchronous-prompt tools (`AskUserQuestion`, plan mode) on headless agents. Set `bypassPermissions` for the old bypass-everything behavior (disposable sandboxes only); other Claude modes (`plan`, `acceptEdits`, `dontAsk`) pass through. Note: `auto` requires Claude Opus 4.6+ / Sonnet 4.6 on the Anthropic API (not Bedrock/Vertex); on other models/providers it is unavailable and Claude Code falls back.
`interfaces.telegram`	map	—	Manager-only. 1:1 Telegram bot for this manager (presence implies it receives Telegram forwards and may call `reply_to_user`).
`autonomy`	string	`low_risk_only`	`full` · `low_risk_only` · `proposal_only`. Policy metadata; runtime permissions and role instructions determine action boundaries.
`can_dm`	list	`[]` = unrestricted	Short-names this agent may DM.
`can_broadcast`	list	`[]` = unrestricted	Channel names this agent may post to.
`reports_to`	string	—	Optional reporting relationship; omit it for peers. Does not itself implement delegation or restrict DMs.
`effort`	string (enum)	wrapper default	Per-agent reasoning effort: `low` · `medium` · `high` · `xhigh` · `max`. Renders as `EFFORT=<value>` and is passed to the runtime (e.g. `claude --effort <value>`). Strict: an unknown value fails validation rather than silently falling back.
`ultracode`	bool	`false`	Per-agent ultracode opt-in. When `true`, emits `"ultracode": true` into the agent’s rendered Claude Code settings (the file passed via `--settings`); the resulting session behavior is Claude Code’s. A Claude Code setting, distinct from the `effort:` row above. Claude Code agents only; on other runtimes it renders nothing and logs a warning.
`on_rate_limit`	list	global default	Overrides the global rate-limit hook chain for this agent. Each entry is the `name` of a hook defined under the global `rate_limits.hooks`. Unset falls back to the global `rate_limits.default_on_hit` chain (which is `[wait]` when empty).
`display_name`	string	agent id	Human-friendly label the TUI shows (roster, detail header, mailbox attribution, statusline) in place of the agent id. Non-empty, ≤64 chars. Render-time only: the agent id stays canonical for routing, tmux session names, and YAML cross-refs (`can_dm`, `can_broadcast`, `reports_to`).
`hooks`	list	`[]`	Per-agent Claude Code hooks, merged additively into the agent’s rendered `settings.json` (the built-in interactive-prompt deny hook keeps precedence). Each entry: `{ event, matcher?, command }`, where `command` is a compose-root-relative path resolved like `role_prompt`. Claude Code agents only; on other runtimes it renders nothing and logs a warning.
`mcps`	map	`{}`	Per-agent MCP servers, merged alongside the built-in `team` mailbox server (a declared server named `team` is rejected). Keyed by server name; each value: `{ command, args?, env? }`. Runtime-agnostic: rendered for any runtime whose descriptor sets `supports_mcp`, skipped with a warning otherwise.
`subagents`	list	`[]`	Per-agent Claude Code sub-agents. Each entry is a compose-root-relative path to a sub-agent markdown file (frontmatter `name` / `description` / optional `tools` / `model`; body becomes its system prompt). Passed via Claude Code’s `--agents`, additively on top of the project `.claude/agents/`. Claude Code agents only; on other runtimes it renders nothing and logs a warning.
`skills`	list	`[]`	Per-agent Claude Code skills. Each entry is a compose-root-relative path to a skill directory (the folder holding `SKILL.md`). Materialized into a per-agent scope dir and exposed via `claude --add-dir`, additively on top of the project `.claude/skills/`. Claude Code agents only; on other runtimes it renders nothing and logs a warning.
`interfaces.telegram.bot_token_env`	string	—	Env var holding the BotFather token. Populated by `teamctl bot setup`.
`interfaces.telegram.chat_ids_env`	string	—	Env var holding a comma-separated allow-list of chat ids.