search

name	search
tier	essential
description	Search the brain's learnings with hybrid BM25+vector recall - filter by project or global, show ranked results (ID/title/snippet/score), and pull a learning into context. Usage: /search <query> [--project <slug> | --global] [--pull <id>] [--limit <N>]
disable-model-invocation	false
allowed-tools	["Read","mcp__igris-brain__igris_memory_hybrid_search","mcp__igris-brain__igris_memory_get"]
triggers	["SEARCH","search the brain","search learnings","find a learning","recall about"]

SEARCH — Universal Brain Search

The interactive recall interface over the brain's learnings. Runs the existing igris_memory_hybrid_search (RRF-fused BM25 + vector) read tool, presents ranked results, and pulls a chosen learning's full content into the session with --pull.

This skill only calls the two MCP tools and formats their output — it does NOT re-implement ranking, weighting, or the BM25-only fallback (the tool owns all of that). It contains zero harness-specific branches.

Arguments

$ARGUMENTS = <query> [--project <slug> | --global] [--pull <id>] [--limit <N>]

Parse as follows:

1. Free-text tokens (everything that is not a flag or a flag value) → the search query.
2. --project <slug> → scope the search to that project (project: <slug>).
3. --global → search everything; omit the project param entirely (broadest scope). Mutually exclusive with --project — if both are given, prefer --global and note it.
4. Neither flag → default to the current project slug (the most useful default). Always print the resolved scope, and hint that --global widens.
5. --pull <id> → skip the search entirely and go straight to the fetch path (§3). If a query is also present, --pull still wins.
6. --limit <N> → pass through to hybrid_search.limit (tool default 10).

Degradation (applies to every brain call)

If the igris-brain MCP server is unavailable, do not block. Warn ONCE: Note: brain MCP unavailable — learnings can't be searched this run. Re-run /search when the brain is reachable. Then stop gracefully. Never error, never block, never emit a stack trace — same convention as /reuse, /scan, and /boot.

1. Search

Call the hybrid search tool with the parsed query and scope:

igris_memory_hybrid_search({ query: "<query>", project: <slug or omitted>, limit: <N or omit> })

Accept the tool defaults — do NOT pass bm25_weight / vector_weight / rrf_k unless the operator explicitly asked to tune them. The tool returns RRF-fused ranked rows, excludes pending_review rows (conscious channel only), and falls back to BM25-only internally when sqlite-vec / embedding is unavailable.

2. Present results

Render a compact table from the tool output plus a scope line:

Scope: project `igris-ai`   (or "Scope: global")

| ID  | Title                                   | Snippet                          | Score |
|-----|-----------------------------------------|----------------------------------|-------|
| 142 | Two-DB drift (fixed in FR-120)          | Pre-FR-120 MCP was http→VPS...   | 0.031 |
| 389 | Tree-aware vendor preserves nesting     | <doc pointer: coding_guidelines> | 0.028 |

Pull full content with /search --pull <id>.

• For rows the tool flagged as promoted-to-doc (FR-200 M2), surface the doc pointer the tool returned as-is — do NOT fabricate a snippet.
• No results → say so plainly and suggest widening with --global or a looser query. Never error.

3. Pull a learning (--pull <id>)

Fetch one learning's full content into the session so it becomes working context:

igris_memory_get({ id: <id> })

Print the full learning — title, category, scope, content, tags, source_brief, confidence — so it lands in the session as working context. If the id does not resolve, say so plainly and offer to re-run the search.

4. Summary hint

After a search, remind the operator of the two follow-up moves:

• /search --pull <id> to read a full learning.
• /search <query> --global to widen scope when a project-scoped search is thin.

---

Constraints

1. Read-only — /search never stores, updates, or deletes a learning. It only reads via igris_memory_hybrid_search and igris_memory_get.
2. Don't re-rank — accept the tool's RRF ordering and defaults; never reimplement weighting or the BM25 fallback in the skill.
3. Explicit scope — always print the resolved Scope: line so the operator knows whether results are project-scoped or global.
4. Graceful degradation — brain absent → warn once, stop, never block.
5. Harness-agnostic — zero harness-specific branches; the body only calls the two MCP tools and formats their harness-agnostic output.