second-brain-query

name

second-brain-query

description

Query the user's Second Brain vault without writing vault Markdown files. Use when the user asks "检查下我的第二大脑", "查旧知识", "第二大脑里有没有", "what do I know about", or asks where a topic should live in the Second Brain. Uses Basic Memory search first with one search-only reindex before Grep fallback, indexes as navigation only, domain-placement mode for read-only routing advice, daily only for explicit journal/timeline evidence, synthesis with wikilink citations for wiki pages, and explicit handoff to ingest for any write or web expansion.

Second Brain Query

Answer from the existing vault. Query is read-only.

This skill is a vault-specific fork of the prior upstream wiki-query framework. Keep strategic retrieval and synthesis; remove hot cache, query filing, auto web search, and writeback.

Execution Contract

Run the retrieval workflow completely before answering. The user should not need to say "follow the skill carefully"; every query request already includes that requirement.

Do not answer from the first search snippet, title, or index entry alone. Read the needed source page bodies, apply path filters, repair Basic Memory once when the local search index is dirty, use Grep fallback only when Basic Memory MCP/CLI cannot be used, and only then synthesize. Stop early only when the query is ambiguous enough to require scope clarification, or when no retrieval channel can run; in those cases, report the blocker instead of giving a partial answer as complete.

In Claude Code, use the AskUserQuestion tool for every scope clarification, "open page vs ingest" decision, or same-topic confirmation. Do not ask these workflow questions as plain assistant text unless the current host does not expose that tool.

Read-Only Boundary

Do not write any vault file.

Do not create pages, fix links, update indexes, append journals, or save answers. If the user wants to keep or expand the result, explicitly hand off to second-brain-ingest.

Query may run basic-memory reindex --project second-brain --search once to repair the local Basic Memory search index before retrieval. This updates Basic Memory's local index, not the Markdown vault, and is allowed inside Query.

Do not use WebSearch/WebFetch inside Query. Web is an ingest source, not a query fallback.

Retrieval Order

Default path:

1. Run basic-memory status --project second-brain --json.
2. If status is dirty, run basic-memory reindex --project second-brain --search once, then run status again.
3. If status is clean, use Basic Memory MCP search_notes with project: second-brain.
4. If MCP is unavailable but CLI search works, use basic-memory tool search-notes "<query>" --project second-brain --page-size <n>; do not rely on the current default project.
5. Filter results by file_path: use wiki/{domain}/{slug}.md content pages as answer sources, _index.md and root index.md as navigation only. Exclude daily/** by default; use daily notes only when the user explicitly asks for journal, timeline, or session-history evidence.
6. Ignore CLAUDE.md, AGENTS.md, GEMINI.md, .claude/**, .agents/**, .gemini/**, .claudian/**, .workflows/**, .brv/**, .obsidian/**, templates/**, raw/**, and deprecated scaffold paths as answer sources. Raw files are provenance/evidence/instructional visual attachments, not query answer nodes.
7. Read the top matching wiki pages needed to answer.
8. Follow only high-value wikilinks from wiki content pages that materially improve the answer. Do not use daily notes as graph traversal sources; when explicitly requested, daily notes are plain file/date evidence only.
9. Synthesize with wikilink citations.

Do not rely on Basic Memory's generic type field for page kind; use file_path.

Fallback when Basic Memory is unavailable or still stale after one reindex:

1. Grep wiki/ and root index.md. Add daily/ only for explicit journal, timeline, or session-history questions.
2. Classify results as exact, related, or none.
3. Do not invent hybrid search scores.
4. Report Grep fallback: <reason> in the final answer.

Daily results are not knowledge pages. When explicitly requested, cite them by plain file path and date only; do not return them as wikilinks, backlinks, or related-note edges.

Raw file paths may be mentioned only as provenance already cited by a wiki content page. Do not traverse from raw files or treat them as related notes.

Treat Basic Memory as stale only after one --search reindex still leaves status dirty, or when search snippets contradict the file on disk. Query remains vault-file read-only: do not edit Markdown to repair anything. Report stale-index residual risk and use Grep fallback only after the Basic Memory MCP/CLI path cannot be trusted.

Index-first path:

Use root index.md -> wiki/{domain}/_index.md -> content page only when the user explicitly asks to start from the index/catalog or when search results are too broad.

Domain-placement path:

Use this read-only mode when the user asks where a pasted topic belongs, whether a new domain is needed, or how to classify a future ingest. Read root index.md and all wiki/*/_index.md, then search candidate pages through the same Basic Memory-first freshness gate. Return candidate domains and likely actions, but do not create or update anything. If the user wants to write, hand off to second-brain-ingest, which must run its own domain-routing preflight before writing.

Domain-placement output should include:

• top 3 candidate domains with reasons,
• any exact or related existing pages,
• whether this looks like update, new page in existing domain, split, ambiguous, or possible new domain,
• what evidence is missing before ingest.

Result Handling

If found:

• Return matching wiki pages as wikilinks.
• Return daily evidence, when explicitly requested, as plain paths/dates rather than wikilinks.
• Explain why each page is relevant.
• State what is still missing or uncertain.
• Ask whether the user wants to open a page or convert the topic into an ingest task.

If ambiguous:

• Show the top candidates.
• Ask the user to confirm the intended scope or whether candidate pages count as the same topic.

If not found:

• State that the vault has no relevant content.
• Ask whether the user wants to ingest this topic.
• End without writing if the user does not switch to ingest.

Completion Criteria

Report blockage instead of answering as complete if any applicable item is missing:

• Query remained read-only.
• Basic Memory freshness gate ran before trusting Basic Memory results.
• If Basic Memory was dirty, one basic-memory reindex --project second-brain --search was attempted before Grep fallback.
• Grep fallback, if used, was reported with the reason.
• Results were filtered by file_path before use.
• Candidate wiki page bodies were opened before synthesis.
• _index.md and root index.md were used only as navigation/scope evidence.
• Daily notes were used only when explicitly requested, and cited as plain paths/dates.
• Raw files were not traversed as answer nodes.
• The final answer states matches, uncertainty, gaps, and whether ingest is needed for any write or web expansion.

Output Shape

Use concise structure:

## 结论
...

## 命中页面
- [[page]] — why it matches

## 缺口
...

<If a follow-up decision is needed in Claude Code, ask it with AskUserQuestion instead of plain text.>

Skip sections that do not apply.

Forbidden Upstream Behavior

Never follow these upstream wiki-query behaviors in this vault:

• Read wiki/hot.md.
• Read deprecated wiki/index.md; query routing must use root index.md instead.
• Save answers under wiki/questions/.
• Append wiki/log.md.
• Automatically WebSearch.
• Turn a query answer into a wiki page without an explicit ingest transition.