메뉴
Decision framework for when and how to delegate work to Retainer's specialist agents. Covers the two specialists in this reference cut (observer, scheduler) plus the cog's direct tools.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
How the observer specialist works in Retainer — role, expected inputs, output shape, quality bar, common pitfalls. The observer reads this at the start of every dispatch.
How the Retainer scheduler specialist works — role, expected inputs, output shape, common pitfalls. The scheduler reads this at the start of every dispatch.
When the Retainer cog should delegate to the observer, how to brief it, what to expect back, and how to verify the result. Observer is the knowledge gateway — recent (hot index) and deep (full archive). Read before dispatching to agent_observer.
When the Retainer cog should delegate to the scheduler, how to brief it, what to expect back, and how to verify the result. Read before dispatching to agent_scheduler.
How dispatching to specialist agents works in Retainer's current shape — the agent_<name> tool shape, what the reply envelope looks like, how errors propagate. Read with `delegation-strategy` (which covers when to delegate).
Recovery patterns in Retainer's current shape for tool errors, agent failures, and cycle abandonment — when to retry, when to narrow, when to escalate. Read after seeing an unexpected error or when planning a multi-step task that might fail partway.
| name | delegation-strategy |
| description | Decision framework for when and how to delegate work to Retainer's specialist agents. Covers the two specialists in this reference cut (observer, scheduler) plus the cog's direct tools. |
| agents | cognitive |
Delegate when the task matches a specialist's scope. Retainer ships two specialists in this reference cut:
| Task Type | Agent | When NOT to delegate |
|---|---|---|
| Any memory-shaped question — recent OR deep. Cycle inspection, recall, fact lookup, CBR retrieval/curation, deep-archive search, cross-store connections, pattern mining | observer | Quick status check the user asked about explicitly — answer directly. Memory tools you can call yourself (memory_read / memory_query_facts) for fact-key-shaped lookups |
| Autonomous-cycle scheduling — cron-recurring or one-shot prompts the cog runs on its own | scheduler | A one-off thing the operator wants run NOW — that's a regular cycle, not a scheduled one |
The observer routes recent vs deep internally based on the question — you don't need to pre-classify.
Web search and URL fetch live on the cog directly via web_search and fetch_url — no separate researcher specialist. Just call them.
Do NOT delegate for:
memory_write, memory_read, memory_clear_key, memory_query_factsagent_observer, agent_schedulerweb_search, fetch_url (on the cog directly — no separate researcher specialist in this cut)read_skill (consult any skill listed in <available_skills>)request_more_turns (extend the cog's tool-turn budget mid-cycle when a deep delegation chain warrants)The memory tools are synchronous and cheap. Use them directly when the answer is "what value did I store under this key?" — not via the observer's get_fact.
Observer (agent_observer) — knowledge gateway. Recent activity (cycle inspection, recall, facts) AND deep archive (full JSONL search, cross-store connections, pattern mining). Also handles CBR retrieval + curation. Never acts on the world. Read agents-using-observer before dispatching.
Scheduler (agent_scheduler) — autonomous-cycle scheduling. Recurring (cron) and one-shot (RFC3339) prompts the cog will run with SourceAutonomous when their time arrives. Read agents-using-scheduler before dispatching.
The per-agent agents-using-* skills cover briefing, expected reply shape, verification, and cost. This skill is the master TOC; the details live there so they can evolve per-agent without rewriting the routing decision tree.
When a specialist returns a summary:
Retainer dispatches to one specialist at a time. If a task genuinely needs multiple perspectives, do them sequentially yourself, or escalate to the operator.
Every delegation is its own cycle of LLM calls (the specialist's react loop runs up to its max_turns). The observer is max_turns=8; the scheduler is max_turns=6. A simple question that costs you one LLM call costs the observer up to 8 if you delegate. Don't delegate "what's 2+2."
If the work has more than 2-3 logical steps, write down what you're going to ask the specialist to do before dispatching. The instruction you pass via agent_<name> is the brief — vague briefs produce vague summaries.