Answer questions from the Registrar's governed corpus and attach a returned qmd:// citation to every
load-bearing claim. Refuse to present unsupported recall as team knowledge.
Overview
The Registrar is the govern layer of Bob's Big Brain: the Compiler prepares candidate knowledge,
deterministic policy decides what becomes durable, and qmd retrieves the governed result. This skill
uses the Registrar-native teamkb_search surface. The unified governed-second-brain plugin exposes a
different tool name, brain_search; do not mix the two contracts.
Prerequisites
- Install the public
intent-brainplugin from this repository for remote, read-only search, or buildapps/mcp-serverfor the full local operator runtime. - For the shipped remote plugin, configure
TEAMKB_API_URLand a per-userTEAMKB_API_TOKEN. - Read the runtime contract for exact inputs, authentication, modes, and failure ambiguity.
Authentication
The shipped remote client sends TEAMKB_API_TOKEN as an Authorization: Bearer header to
TEAMKB_API_URL. Supply it through the plugin environment and never print, capture, or commit it. A
locally built Registrar server can search a local index without API authentication.
Instructions
Step 1: Search curated memory
Derive 1–4 distinctive keywords from the question, dropping generic question words. Call
teamkb_search with scope: "curated" and an optional limit from 1 through 50.
teamkb_search({ query: "Caddy reverse proxy", scope: "curated", limit: 10 })
If the result is empty, retry the same keywords once with scope: "all". Use inbox, archived, or
bulk only when the user explicitly requests that uncurated, retired, or bulk-digestion material.
Step 2: Answer only from results
- Synthesize a short answer from returned snippets.
- Attach the exact returned
qmd://citation to each supported claim. - Surface conflicting results with both citations.
- Label reasoning beyond the cited snippets as inference.
- Never invent a citation or silently fill a gap with general knowledge.
Step 3: Handle no evidence
After the curated-to-all retry, state plainly that no governed evidence was found. The legacy remote client maps several transport and authorization failures to an empty result, so do not claim that an empty response proves the corpus lacks the fact; mention this limitation when the miss is unexpected.
Output
- Direct answer or honest no-evidence statement.
- Inline
qmd://citation after every load-bearing claim. - A Sources list containing each distinct citation used.
Examples
/brain what does the system map say about the reverse proxy?
The reverse proxy is the single ingress (qmd://kb-curated/9c2e42f1-7b60-4ed2-a9dd-648d6c786d43.md).
Sources:
- qmd://kb-curated/9c2e42f1-7b60-4ed2-a9dd-648d6c786d43.md
If both scopes return no results:
No governed evidence was found for that topic. The legacy remote client can also return an empty set
when the service or credential fails, so an operator should check connectivity if this miss is
unexpected.
Error Handling
| Situation | Response |
|---|---|
source is unconfigured |
Configure TEAMKB_API_URL; do not describe this as an empty corpus. |
Empty brain-api result |
Complete the scope retry, then state no evidence was returned and note the legacy ambiguity when relevant. |
| Tool is unavailable | Enable the Registrar intent-brain plugin or locally built teamkb MCP server. |
| User asks to write | This skill is read-only; use the Registrar operator brain-save skill only with the full local server. |
Guardrails
- Treat inbox, archive, and bulk results as explicitly requested context, not curated truth.
- Provenance identifies where a capture came from; it does not prove the content is true.
- Prefer a narrow cited answer over a broad unsupported one.
Resources
- Bob's Big Brain Registrar
- Bob's Big Brain umbrella
- tobi/qmd, the attributed retrieval engine