agent-switchboard-ops

Start a managed Claude supervisor with start_managed_claude_supervisor, providing a specific objective, policy with red lines, and stall_timeout_seconds. Then dispatch bounded task slices via send_to_managed_claude_session, one verifiable increment at a time, and track progress with wait_supervisor_event long polling.

Same-vendor execution work defaults to the host's native subagent mechanism, while cross-vendor review, debate, or consultation routes through the switchboard broker via route_agent_task or consult_* tools. Small single-session tasks skip delegation entirely, and downgrade only happens after a real call fails.

Worker refusals have two causes: the session is restricted to read-only advice mode, or the dispatch prompt lacks an explicit authorization statement. Fix the first by checking permission mode and routing; fix the second by prepending an authorization declaration stating the worker is explicitly authorized to read code, run scripts, and create files.

Yes, the Skill still triggers through the host's native subagent channel when the broker is unavailable, but you lose the broker's event stream, archiving, and cross-session continuation. Never guess tool names or conclude a tool is unusable from the declared catalog alone; verify with a real call first.

Supervisors hang when closure relies on the manager remembering to call close_supervisor. The Skill mandates setting stall_timeout_seconds at startup, running a list_managed_claude_supervisors self-check at each round end, and closing or stopping any supervisor whose task is finished or whose event stream has stalled.

Use a logged-in CLI worker as the search channel: queue a codex_cli request with an explicit cheap model tier like gpt-5.6-luna and a structured prompt requiring source URLs, then collect results with a single request_result long poll. This reuses existing subscriptions at zero added cost with 1-3 minute latency.