search-summaries — L1 retrieval over L2 structures

Framework context

L1 retrieval skill whose substrate is the L2 summary table. Each result IS a citation back to the underlying conversation (via conversation_id) and to the summary row (for downstream L3 consumers). The "extract mode" parameter projects different structured fields from the summary payload — still no synthesis, just projection.

When to invoke

• User wants decisions, open-questions, concepts, or quotables — not prose conversations
• After semantic-search finds candidate messages and you want the structured version
• Filtered queries: specific domain + importance + stage combination
• extract="questions" → open questions only. extract="decisions" → decisions only. extract="quotes" → quotables only. Default extract="summary".

Input

query: str
extract: "summary" | "questions" | "decisions" | "quotes" = "summary"
limit: int = 10
domain: str | None
importance: "breakthrough" | "significant" | "routine" | None
thinking_stage: "exploring" | "crystallizing" | "refining" | "executing" | None
source: str | None
mode: "hybrid" | "fts" | "vector" = "hybrid"

Output contract

Markdown with per-row citations. Every row carries [conv_id · date] and [summary_id]. Extract modes project different structured fields but never paraphrase.

Does NOT do

• Synthesize across summaries (that's L2 what-do-i-think)
• Fabricate filter values — invalid enum values return empty with explicit error
• Bypass sanitization (WHERE-clause allowlist [a-zA-Z0-9\s\-_.,] is enforced)
• Combine extract modes — one projection per call

Verification checklist

• [ ] Every row has conv_id + summary_id citation
• [ ] Filter values pass sanitizer before SQL
• [ ] Mode fallback is explicit if LanceDB unavailable
• [ ] Re-run with identical args → identical order