kb-find（知识库查找）

/kb-find

Progressive 4-tier KB discovery. Read-only.

Constraints

• NEVER Read index.md or config.json — kb_loader handles these; config auto-detects.
• Combine reasoning + tool calls per turn. Batch parallel Bash/Read calls.
• Only Read individual concept .md files (Tier 2/3/4).
• Global KBs (shown with [read-only] badge in topic listing) are searchable but not writable. Treat them identically for discovery.

1. Input Parsing

Parse query from $0, strip flags. Two modes:

• Default: find concept notes matching query keywords.
• --challenge: find content that contradicts or complicates the query. Score by contradiction potential. Include adjacent topics. Read 1-2 extra notes in Tier 3/4. Skip KB meta-topics unless query is about them.

2. Tier 1 — KB Loader Queries

KB_LOADER="${CLAUDE_SKILL_DIR}/scripts/kb_loader.py"

1. $KB_LOADER --list-topics — tree by @kb-name with [skill] tags.
2. Pick relevant topics. Slash paths (compact) or @kb-name/path. Challenge: also contradicting topics.
3. $KB_LOADER --topic <path> per topic. All calls in one turn.
4. Build topic map: {topic_name, notes[], children[], has_skill}.

--topic returns JSON (names/descriptions) for Tier 2 scoring. Empty list → report and stop.

3. Tier 2 — Frontmatter Scan

Skip if Tier 1 name+description clearly shows relevance → Tier 3/4. Only scan ambiguous notes.

1. Read first 15 lines. Extract name, description, summary.
2. Score relevance; promote if warranted. Cap 2-3 candidates.

4. Tier 3 — TOC Section Read

Skip to Tier 4 if no TOC or note is <100 lines.

1. Read first 30 lines (frontmatter + TOC). If already read 15, use offset=16, limit=15.
2. Read relevant sections ±1 adjacent.
3. Follow relevant markdown links → Tier 2. Cross-KB links (@kb-name/path) → --topic @kb-name/topic.
4. Follow soft references (lines starting with see: followed by @kb/topic) → --topic @kb/topic. These are semantic hints to global KBs.

5. Tier 4 — Full Read

1. Full read when no TOC or Tier 3 insufficient. Use offset to skip prior lines.
2. Cap 500 lines/note; note truncation. >10 notes → ask user to narrow.
3. Follow relevant links → Tier 2.

6. Output

Return: ## Topic Map (paths), ## Relevant Concept Notes (path + title + content, label concept vs skill), ## Notes (caveats).

Challenge mode output: ## Topic Map, ## Counter-Evidence (path, strength, type, excerpt), ## Summary (1-3 sentences), ## Notes.