Brain — Query

Single responsibility

Answer a question against the brain wiki, citing source files. Log
unanswered or low-confidence queries as gaps for brain-ingest to
address.

This skill is invoked first by every other skill, per ADR 010.

Scope routing

Scope	What to search
forge-dev	brain/forge-dev/themes/ + category indexes
cycles	brain/cycles/themes/ + category indexes
project	brain/projects/<name>/themes/ + profile.md (forge repo, ADR 035)
all	union of all three (emit a scope-missing warning)

Role defaults (the calling skill or orchestrator should supply these; who reads which brain is CLAUDE.md's brain-read policy, ADR 010):

Role	Default scope
architect / PM	cycles,project
reflector	all
dev-loop / reviewer	project
forge-dev session (no cycle)	forge-dev,cycles

Inputs

• A natural-language question or list of questions.
• Optional: scope — which brain(s) to search (see table above).
• Optional: project name — required when scope=project.
• Optional: category filter (pattern | antipattern | decision | operation | reference).

Outputs

{
  answers: Array<{
    question: string;
    answer: string;             // synthesised answer
    confidence: 'high' | 'medium' | 'low';
    sources: string[];          // brain file paths (theme pages only — not raw, not category indexes)
    gap?: boolean;              // true if confidence is low or no source found
  }>;
}

For each gap: true answer, append to _logs/<cycle-id>/brain-gaps.jsonl.

Event-log entries to emit

• brain-query.start — with the questions.
• brain-query.hit — one event per question that found high/medium-confidence sources.
• brain-query.gap — one event per question with low/no confidence.
• brain-query.end — summary.

Process

Keyword scan (all scopes)

1. Read the relevant category index for the scope:
   • Brain 1 (forge-dev): brain/forge-dev/decisions.md or brain/forge-dev/reference.md
   • Brain 2 (cycles): brain/cycles/patterns.md, antipatterns.md, operations.md, or decisions.md
   • Brain 3 (project): brain/profile.md + brain/themes/ listing
2. Read 2–5 theme files whose slugs or one-liners match the question keywords.
3. Synthesise + cite. Write a one-paragraph answer preserving exact terminology from the cited themes. Cite by file path. Score confidence:
   • High: ≥ 2 corroborating themes, all on-topic.
   • Medium: 1 source on-topic.
   • Low / gap: no good source — set gap: true.
4. If still nothing after reading the category index + 2–3 theme candidates, mark gap: true.

Constraints

• Cite, don't paraphrase deeply. Synthesis is a one-paragraph answer + source list. The caller can read the linked file.
• Cite theme pages only. Valid sources entries are brain/cycles/themes/<slug>.md, brain/forge-dev/{log.md,decisions.md,reference.md}, and brain/projects/<name>/{profile.md,themes/<slug>.md}. Never cite brain/cycles/_raw/* (inputs to synthesis, not citations) or category indexes (navigation, not knowledge).
• Exhaustive on theme coverage. Recall matters — under-citing (missing the corrective antipattern) is worse than over-citing by 1–2 extras.
• Scope is load-bearing. When scope: project, cited sources MUST come from inside the project's own brain. Do not pull in forge-dev concepts not documented in the project brain — that's hallucination.
• Missing scope defaults to all + warn. Emit: [brain-query] no scope supplied — searching all three brains; supply --scope to reduce noise.
• Gaps are logged, not silently failed. If the brain doesn't know, the brain learns by the next ingest pass. Naming-the-absence without gap: true is the worst failure mode.
• No web fallback in this skill. Broader research is the calling skill's responsibility.
• Fast model by default. Haiku is the default; per-skill override via the calling skill's frontmatter.