| name | search-history |
| description | Use when the user asks about prior context in this cairn ("was X considered?", "what did we decide about Y?", "has anyone looked at Z?"). Searches local files — meetings, findings, decisions, open questions — without relying on a vector index or external service. Returns excerpts with source path, date, and author so answers can be cited. |
Search prior discussions
The cairn's history lives in plain files on disk. For most queries, that's all you need.
Where to search
state/decisions.yaml — keyword scan; each entry has id, date, author, decision, context.
state/open_questions.yaml — keyword scan; each entry has id, raised_by, date, question.
knowledge/meetings/*.md — one file per meeting, named YYYY-MM-DD.md. Speaker-attributed.
knowledge/findings/*.md — logged findings with frontmatter (date, author).
knowledge/literature/ — paper notes.
How to search
-
Identify keywords from the user's question. Include obvious synonyms but don't expand too aggressively — exact-string matches are usually what the user wants.
-
Use grep -rinE (or the editor's equivalent) across state/, knowledge/, and branches/. Read excerpts in context (a few lines before and after the match), not just the matched line.
-
For YAML state files, prefer reading the whole entry that contains the match — a decision is small, and the context field often carries the actual reasoning.
-
Filter by branch. If the user is on a non-main branch, scope to the branch's view of state files (git show <branch>:state/decisions.yaml). Don't accidentally cite something that was decided on a branch the user has never seen.
How to answer
Cite each excerpt with source path + date + author. Examples:
- "On 2026-04-22, kyle recorded D-014: Use stratified resampling… (context: SMOTE was the alternative)."
- "Maria raised Q-012 on 2026-05-08 about whether the bias correction introduces identifiability problems — still open."
- "In the 2026-05-15 meeting (
knowledge/meetings/2026-05-15.md), the group discussed…"
Empty results: say so plainly. Don't fabricate.
When to escalate beyond local files
This skill is intentionally low-tech. If the cairn has an MCP server attached and the user's question is genuinely semantic (no obvious keywords), prefer the MCP tool find_related_prior_discussion. Otherwise, file scans are faster and avoid invented relevance.
Acceptance criteria (US-A-05)
- No embeddings or external services required.
- Each result is attributable (source path, date, author).
- Results are scoped to the current branch's view of the cairn.