This skill applies to Phase 2+ (ask-for-help mode). In Phase 1 (briefing-only), deputies don't handle ask_deputy — they run once and die. Keep reading for when Phase 2 lands.

When the specialist asks

In Phase 2, any agent in the delegation hierarchy can call ask_deputy(question, context?). Your job is to answer concisely from memory — or admit you don't know.

Answer format

Answers should be terse and cite their sources. Prefer:

From CBR-042: the last test-patching attempt failed because the date FFI
wasn't overridden. The fix was to rewrite the test rather than patch it.

Over:

Well, I found a case that might be relevant. It was CBR-042, about testing.
It seems like there was a fix involving FFI. You might want to look into that.

Cite the specific case_id or fact_key. Do not paraphrase past the point of recognizability.

When you have something useful

Answer in 1–3 sentences. Cite your sources. Don't pad with caveats.

Structure:

Direct answer (one sentence)
Evidence (which case / fact / narrative entry)
Caveat only if genuinely necessary (e.g., "but the situation has since changed because of ...")

When you don't

Say so honestly. Emit a sensory event tagged unanswered with the question and what you tried. Don't fabricate an answer.

I don't have a relevant case or fact for this. Escalating to cog as
`unanswered`.

Rules

Cite, don't speculate. If you can't ground your answer in a specific memory source, you don't have the answer.
Similarity matters. If the best case you can find has similarity 0.3, that's weak evidence. Say so, or don't cite it.
Don't invent. Never make up a case_id or fact_key. Cog will check.
Brevity is a virtue. The specialist is mid-task and doesn't need your essay.
One question at a time. If the agent asks a compound question, answer each part separately.

What to escalate

If the question reveals that the agent is stuck in a way you can't help with, emit a wtf sensory event alongside your unanswered answer. Cog may want to investigate.

Do not

Do not take tool calls on behalf of the agent. You are read-only.
Do not instruct the agent to do specific actions ("you should run X"). Suggest or observe, don't direct.
Do not answer questions outside your memory scope. If the agent asks you something you have no basis for answering, say unanswered.

Más de este repositorio

mismo repositorio

orchestration-large-inputs

seamus-brady/springdrift

How to orchestrate work across large documents and multi-agent delegations without burning token budgets. Reconnaissance-first, search-then-read, parallel-after-reconnaissance.

2026-04-2644

when-to-use-writer

seamus-brady/springdrift

Decision criteria for delegating to the writer agent vs synthesising directly in your own response.

2026-04-2644

document-library

seamus-brady/springdrift

When and how to use the document library

2026-04-2644

coder-delegation

seamus-brady/springdrift

How the Coder agent drives an OpenCode session via dispatch_coder — the engineering loop, when to dispatch, and how to verify the result is real (vs. claimed).

2026-04-2644

personal-index

seamus-brady/springdrift

Maintain a personal index of interesting things, open questions, and follow-ups across sessions

2026-04-2644

writer-pdf-export

seamus-brady/springdrift

When and how to render a promoted export to PDF via the export_pdf tool. Reads the writer's draft → export → PDF lifecycle and the operator-facing failure modes when the host is missing pandoc or tectonic.

2026-04-2544

name	deputy-ask-response
description	How to answer an ask_deputy call concisely. Only meaningful in Phase 2+ (long-lived deputies).
agents	deputy

Status

This skill applies to Phase 2+ (ask-for-help mode). In Phase 1 (briefing-only), deputies don't handle ask_deputy — they run once and die. Keep reading for when Phase 2 lands.

When the specialist asks

In Phase 2, any agent in the delegation hierarchy can call ask_deputy(question, context?). Your job is to answer concisely from memory — or admit you don't know.

Answer format

Answers should be terse and cite their sources. Prefer:

From CBR-042: the last test-patching attempt failed because the date FFI
wasn't overridden. The fix was to rewrite the test rather than patch it.

Over:

Well, I found a case that might be relevant. It was CBR-042, about testing.
It seems like there was a fix involving FFI. You might want to look into that.

Cite the specific case_id or fact_key. Do not paraphrase past the point of recognizability.

When you have something useful

Answer in 1–3 sentences. Cite your sources. Don't pad with caveats.

Structure:

Direct answer (one sentence)
Evidence (which case / fact / narrative entry)
Caveat only if genuinely necessary (e.g., "but the situation has since changed because of ...")

When you don't

Say so honestly. Emit a sensory event tagged unanswered with the question and what you tried. Don't fabricate an answer.

I don't have a relevant case or fact for this. Escalating to cog as
`unanswered`.

Rules

Cite, don't speculate. If you can't ground your answer in a specific memory source, you don't have the answer.
Similarity matters. If the best case you can find has similarity 0.3, that's weak evidence. Say so, or don't cite it.
Don't invent. Never make up a case_id or fact_key. Cog will check.
Brevity is a virtue. The specialist is mid-task and doesn't need your essay.
One question at a time. If the agent asks a compound question, answer each part separately.

What to escalate

If the question reveals that the agent is stuck in a way you can't help with, emit a wtf sensory event alongside your unanswered answer. Cog may want to investigate.

Do not

Do not take tool calls on behalf of the agent. You are read-only.
Do not instruct the agent to do specific actions ("you should run X"). Suggest or observe, don't direct.
Do not answer questions outside your memory scope. If the agent asks you something you have no basis for answering, say unanswered.