searchlisted

# sessions:search Recall is not "run one CLI query and paste the result." The index that backs `agents sessions "<q>"` only covers **user turns + title/topic/project** (`session_text`, bm25) and, separately, **tool activity** (`tool_call_text`, trigram, needs `--include tools --query`). **Assistant answers are never indexed.** A session can be 665 messages long with only a 43-character title in the index — the actual answer the user wants is buried in an assistant turn the index cannot see. Follow this procedure in order; do not skip straight to guessing an answer from memory. ## 1. Derive search terms from the user's ask Pull out, in this priority order, every one of these that appears in the ask: - **Ticket/issue IDs** — `PHNX-\d+`, `RUSH-\d+`, `#\d+`, or any tracker key shape. - **Filenames / paths** — anything with a `/` or a file extension. - **Symbols** — function/class/variable names, CLI flag names, config keys. - **Distinctive keywords** — 2-4 words that would be rare together (skip generic verbs like "fix", "add", "look at"). Keep 2-5 terms. Too many terms narrows an FTS query to zero; too few floods it. If the ask is vague ("what did we decide about the recall design"), use the closest concrete nouns ("recall design") — do not ask the user to rephrase before trying at least one layer below. ## 2. Layered discovery — try each in order, stop at the first that returns real hits **Layer 1 — user turns + titles (fastest, whole-session refs):** ```bash agents s