iterative-retrievallisted

@markdownai v1.0 # Iterative Retrieval Use when looking for context about a topic before reading files or writing code. ## Retrieval Ladder (run in order — stop when you have enough) **Rung 1 — KB search** (broadest, fastest): Call `kb_search(query="<topic>", fields=["id","title","summary"])`. Read titles and summaries. If 2+ relevant atoms found, go to Rung 1b. **Rung 1b — KB get** (precise, low-token): If you already have an atom id, call `kb_get(id="<ATOM_ID>")` (embeddings are omitted by default). **Rung 2 — Store search** (collection-scoped): Call `soil_search(collection="hanuman/atoms", query="<topic>")` or `soil_search` on `hanuman/file-index`. Use when KB search returns nothing. **Rung 3 — Temporal query** (if currency matters): Call `kb_at(query="<topic>", at_time="<ISO>")` to get KB state at a specific point in time. **Rung 4 — JELES retrieval** (session history): Call `mem_jeles_extract(app_id, query="<topic>")` to pull from indexed session JSONLs. **Rung 5 — File read** (last resort): Only read the specific file section if Rungs 1–4 returned nothing useful. Read the specific section, not the whole file. ## Rule Never skip to Rung 5. The KB is the map. Files are the territory. Read the map first.