CAO Session Management

Overview

CAO runs multi-agent workflows in named sessions. A conductor agent inside each session orchestrates the work.

Core Concepts

• Session: A group of agent terminals working together
• Conductor: The supervisor terminal — receives instructions, delegates to workers
• Provider: LLM backend. Default kiro_cli, override with --provider

Prerequisites

Before launching a session, verify:

• cao-server is running at localhost:9889. Quick check:

  curl -sf http://localhost:9889/sessions >/dev/null && echo OK || echo "start cao-server"
  

  If not running, start it in a separate terminal: cao-server.
• The agent profile is installed. cao launch --agents <profile> fails if the profile is unknown. Install built-ins or custom files with cao install <profile|path|url>.

Discovering Available Profiles

Profiles are CAO-level entities, installed with cao install regardless of which CLI provider runs them. To find available profiles:

The HTTP endpoint is the recommended check: it scans the built-in packaged store, the local store (agent-store/), and provider-specific directories (including agent-context/), then returns a deduplicated list (by profile name, built-in wins) with a source label on each entry.

If unsure which profile to use, ask the user rather than guessing.

Quick Example

A complete, copy-pasteable supervisor launch. The default provider is kiro_cli; pass --provider <name> to use another (claude_code, codex, antigravity_cli, kimi_cli, copilot_cli, opencode_cli, cursor_cli).

This example assumes a configured CAO setup (server running, profiles installed). On an already-configured host you can skip straight to cao launch. The cao install lines below are only for first-time setup; remove them if your CAO is already configured.

# Optional — skip if your CAO is already configured with these profiles.
# Provider-agnostic: `cao install` works for any provider.
cao install code_supervisor
cao install developer
cao install reviewer

# Launch headlessly (assumes cao-server is already running)
cao launch --agents code_supervisor --headless --yolo \
  --session-name my-task --working-directory '/path/to/project' \
  "Build a hello-world Python script. Delegate to developer, then reviewer."

# Same launch on a different provider
# cao launch --agents code_supervisor --provider claude_code --headless --yolo \
#   --session-name my-task --working-directory '/path/to/project' "..."

# Check progress / final output
cao session status cao-my-task
cao session status cao-my-task --workers

# Clean up
cao shutdown --session cao-my-task

Launching a Session

Every cao launch MUST include:

• --agents PROFILE — see Discovering Available Profiles above; if unclear, ask the user
• --headless — required from an LLM agent; without it cao tries to attach tmux
• --session-name NAME — cao adds cao- prefix automatically
• --working-directory DIR — a wrong path silently breaks the session with no recovery short of shutdown and relaunch. Ask the user if unclear. Always wrap in single quotes to pass the literal path to the server (prevents local shell expansion of ~ or variables before the value reaches cao).

cao launch --agents <profile> --headless --yolo \
  --session-name <name> --working-directory '<path>' "<task>"

--yolo skips confirmation prompts. Required when launching from an agent — interactive prompts will stall the session.

For SOP-driven workflows (Kiro provider): launch with /prompts to discover available SOPs, then send the matched SOP name prefixed with @ (e.g., @my-sop-name), then send the task — each as separate messages after polling for completed status.

Commands

cao session send waits for completion and returns output inline by default. With --async, it sends and returns immediately without waiting. With --timeout N, it waits up to N seconds — if the timeout expires, the agent is still running; check status later. Session names in commands use the cao- prefixed form (e.g. --session-name mywork → use cao-mywork).

Worker Communication

Inside a session, the conductor talks to workers via two MCP tools:

Prefer communicating through the conductor (cao session send SESSION "msg") rather than directly to worker terminals. Bypassing the conductor leaves it without state on what was asked or answered, which causes confusion. Two exceptions: unblocking a stuck worker, and follow-up questions to a persistent async worker (see below).

handoff (blocking) — conductor sends task and waits for the worker to reach COMPLETED status, then reads the output. If it times out, the worker is still running — the conductor just stopped waiting.

assign (non-blocking) — conductor sends task and returns immediately. The worker is expected to call send_message back to the conductor's terminal ID when done. Each terminal only knows its own ID via $CAO_TERMINAL_ID.

By default the conductor uses sync (handoff). You can override this by explicitly asking it to use async or sync protocol when sending it a task.

Async workers (assign) stay alive after completing their task and can answer follow-up questions — useful for ongoing investigation where you want to keep querying the same worker. In this case, sending directly to the worker terminal is appropriate:

cao session send SESSION "<follow-up question>" --terminal <worker-terminal-id>

Common Mistakes

Wrong working directory — agents won't find files, builds fail with confusing errors.

Stuck conductor — conductor is waiting on a worker that stopped responding. Check the worker's status first, then decide: prompt it to continue and send results back, or ask it to resend if it already finished. Never re-delegate work that may still be running — it risks duplicate work.

cao session status SESSION --workers
cao session status SESSION --terminal <worker-terminal-id>
cao session send SESSION "continue your work, then send results to terminal <conductor-terminal-id>" --terminal <worker-terminal-id>

Get the conductor's terminal ID from cao session status SESSION --json.