--- name: cao-supervisor-protocols description: Coordinate, delegate, and split work across CAO worker agents as a supervisor — choose between assign (async fire-and-forget) vs handoff (sync wait-for-result), fan out tasks to parallel specialists, manage idle-based inbox delivery, or debug why worker callbacks aren't arriving. Covers dispatch patterns, result collection, and multi-agent orchestration planning. Not for worker-side task execution, creating skills, managing plugins, or session lifecycle (cao launch/shutdown). compatibility: Requires cao-server running at localhost:9889 and cao-mcp-server configured on the agent. --- # CAO Supervisor Protocols Use this skill when supervising worker agents through CLI Agent Orchestrator. This skill covers how supervisors should dispatch work, decide between `assign` and `handoff`, and receive worker results without blocking inbox delivery. ## Core MCP Tools From `cao-mcp-server`, supervisors orchestrate work with: - `assign(agent_profile, message)` for asynchronous work that returns immediately - `handoff(agent_profile, message)` for synchronous work that blocks until the worker finishes - `send_message(message, receiver_id=None)` for direct messages — `receiver_id` defaults to the terminal that created yours via handoff/assign - `answer_user_prompt(terminal_id, answer)` for answering a Hermes worker that reports `waiting_user_answer` Your own terminal ID is available in the `CAO_TERMINAL_ID` environment variable. CAO appends it to assigned task messages and records it on worker terminals automatically, so you rarely need to handle it yourself. ## Do It Directly or Delegate Delegating has a fixed cost: routing the message, starting a worker, and the worker's own verification. Choose by the shape of the task: - Do a small, fully specified change yourself when you can run its checks. Example: a one-line dependency floor plus a relock. - Delegate when parallel work saves wall-clock time, when the task needs a specialist profile, or when an independent review is worth the wait. Your agent profile's rules take priority over this heuristic: `code_supervisor`, for instance, always delegates coding and never writes code itself. Measured on [#1119](https://github.com/awslabs/cli-agent-orchestrator/pull/1119): the same dependency-floor change took 1m 18s done directly and about 2m 57s routed through a conductor to a developer seat. Both produced byte-identical lockfiles. The delegated result had the better code comment, which is the value an independent pass adds. ## Gotchas | Mistake | Fix | |---------|-----| | Keeping the terminal busy after `assign` — polling in a loop, or running `sleep`/`echo` to wait. | Inbox messages are delivered when your terminal goes **idle**. Dispatch everything first, then end the turn. | | Answering a Hermes worker that reports `waiting_user_answer` with `assign`, `handoff` or `send_message`. | Use `answer_user_prompt(terminal_id, answer)` with the exact selection to submit, such as `1` or `o`. | | Doing a small coding task yourself because the do-or-delegate heuristic allows it, on a profile that forbids it. | Your agent profile's rules win. `code_supervisor`, for instance, always delegates coding. | | Treating a `send_message` callback as proof the work landed, when the worker wrote files. | Ask workers for the deliverable up front, and for **absolute paths** in the callback message. | | Assuming you must hand-write callback instructions into every assigned message. | CAO appends your terminal ID and records you as the worker's caller automatically. An explicit callback ID is optional emphasis; it is required only when `CAO_ENABLE_SENDER_ID_INJECTION=false`. | ## Choosing Between Assign and Handoff Use `assign` when the worker should continue independently and report back later. This is the normal pattern for fan-out work or parallel execution. Use `handoff` when the next step is blocked on the worker result. The orchestrator waits for completion, captures the worker output, and returns it directly to the supervisor. Typical pattern: - Use `assign` for analysis, research, or code changes that can run in parallel. - Use `handoff` for report generation, blocking review steps, or any task where you need the result before you can continue. ## Idle-Based Message Delivery Assigned workers usually return results through `send_message`. Those inbox messages are delivered to the supervisor automatically when the supervisor terminal becomes idle. This means supervisors should: - Dispatch all planned worker tasks first - Finish the turn after dispatching work - Avoid running placeholder shell commands just to wait Do not keep the terminal busy with `sleep`, `echo`, or similar commands while waiting. A busy terminal delays inbox delivery. If you need multiple worker results, dispatch them all first, then end the turn. Do not poll manually in a loop. ## Callback Pattern By default, CAO appends your terminal ID and callback instructions to every assigned message automatically, and records your terminal as the worker's caller — workers can reply with `send_message` without a `receiver_id`. You do not need to hand-write callback instructions. You may still include an explicit callback ID in the task message for emphasis: ```text Analyze dataset A. Send results back to terminal abc123 using send_message. ``` If your deployment disables the automatic suffix (`CAO_ENABLE_SENDER_ID_INJECTION=false`), the explicit pattern above is required: the structural caller record still works, but the worker gets no in-message reminder. ## Direct Supervisor Communication Use `send_message` when you need to contact an existing terminal directly rather than spawning a new worker. Examples: - Relay follow-up instructions to a worker you already created. - Forward a worker result to another coordinator terminal. - Send a concise status update to a collaborating supervisor. When sending direct messages, include enough context that the receiver can act without re-reading the full original task. ## Interactive Worker Prompts Hermes workers can stop on approval prompts or clarify pickers and report `waiting_user_answer`. When a Hermes worker is in that state, do not use `assign`, `handoff`, or `send_message` to answer it. Use `answer_user_prompt(terminal_id, answer)` with the exact selection or text to submit, such as `1`, `o`, or a custom answer. Other providers may still emit prompts in their terminal output without reporting `waiting_user_answer`. For those providers, treat the prompt as ordinary terminal output and answer it with `send_message` or direct input according to the workflow you are running. ## Practical Workflow 1. Dispatch asynchronous workers with `assign` — callback routing is automatic (your terminal ID is appended to the message and recorded as the worker's caller). 2. Use `handoff` only for steps that must finish before you can continue. 3. End the turn so asynchronous worker messages can be delivered. 4. When messages arrive, synthesize the results and continue the workflow. ## Reliability Guidelines - Tell workers exactly what deliverable they should return. - When workers create files, ask them to return absolute paths in their callback message. - Do not assume results will be delivered while your terminal is still busy. - Keep orchestration instructions separate from domain requirements so workers can parse both cleanly.