Assess how much machinery a task deserves, then return either 'handle it yourself', one specialist's system prompt, or a multi-stage plan (planner → executor → tester → reviewer) with intake questions to ask first. Call this before starting any non-trivial task. Respect a 'handle_directly' verdict — spawning a subagent for small work wastes tokens and time.

Fetch the system prompt for one stage of a plan returned by brief. Call it when the previous stage is finished, passing that stage's output as context so this one is briefed on what it is receiving.

Show catalog entries — agents, skills, knowledge, runbooks, topics, tools, connectors — each tagged base, custom, or base+custom.

Fetch the full content of one catalog entry — an agent's prompt, a skill, a runbook with its table of contents, a connector's guardrails.

Retrieve passages from the organisation's own documentation. Use for a direct factual lookup. Cite the returned chunk id when you use a passage, and say so plainly when the answer is not in the corpus rather than filling the gap.

Record a preference, rule, or correction so future briefings honour it. Restating an existing preference reinforces it rather than duplicating; a constraint or correction that contradicts a softer memory retires the old one. Use for standing rules, never for one-off task instructions.

Search remembered preferences, ranked by relevance and how binding they are.

Permanently delete one memory by id. Use when the user says a preference is wrong or obsolete.

Tell the router whether the agent it chose was right. This adjusts the term-to-agent weights it consults, so routing improves with use. Call after the user reacts to a subagent's work.

What is loaded: catalog counts by kind and layer, knowledge index size, stored preferences, and any files that failed to parse.