search_conversations only indexes user prompts, missing work by what was done

[RyanSeanPhillips]

Problem

The search_conversations MCP tool indexes only the prompts the user typed — grouped by session, ranked by recency, with all space-separated terms ANDed. It does not index:

• assistant responses
• tool calls / results
• edited or created file paths
• any per-session summary

Consequence

You can't find a conversation by what was done unless the user happened to type those exact words. Recall fails whenever a later description differs from the original phrasing, and adding search terms makes it worse (AND → zero hits).

Concrete miss

Searching for the phillips-website dark→light theme conversion returned 0 results for every natural query — light theme website, phillips-website theme, site light accent. The session was only relocated via an originSessionId field stamped inside a memory file, not via search. Today the reliable way to relocate past work is memory-file back-links, not the search tool.

Other rough edges

• Ranked by recency, not relevance (no score / BM25).
• No project filter param — must eyeball the project field.
• Strict AND on terms: single-word queries flood, multi-word queries starve.

Proposed improvements (rough priority)

1. Index more than prompts. Pull assistant text + tool_use names + touched file paths from the JSONL so queries match the work, not just the ask. Biggest win for the "what did I work on" use case.
2. Per-session summary index. Generate/store a one-line "what happened" per session and search that — cheaper to scan, robust to phrasing drift.
3. Add a project filter param and relevance ranking; consider OR/fuzzy so extra terms don't zero out results.
4. Surface originSessionId back-links so a topic match in memory can point straight at the source conversation.

Where

The tool lives in the cldctrl MCP server's search_conversations handler, which currently scans JSONL for user-role messages only. Extend the index builder to also emit assistant text + tool names + touched file paths (and optionally a summary line); add optional project and relevance params to the tool schema.

[RyanSeanPhillips]

Fixed in 1d619d2.

What changed: search now indexes the full conversation content — user + assistant text, tool names, and touched file paths (tool results skipped as huge/noisy) — instead of only history.jsonl prompts. Built into a per-session doc cached by (mtime,size) and persisted to search-index.json (same pattern as usage-buckets.json); only changed sessions re-parse. Cold build over all projects ~0.9s, index ~900KB.

Addresses:

1. ✅ Indexes more than prompts — matches what was done. The phillips-website dark→light theme work now surfaces top-ranked for light theme website, phillips-website theme, site light accent (all were 0 before).
2. ✅ Per-session content doubles as the "what happened" index (a one-line gist also feeds session titles separately).
3. ✅ project filter added to searchConversations + the search_conversations MCP tool; relevance ranking (coverage → occurrences → recency) replaces pure recency; terms are ranked-OR'd so extra words refine instead of zeroing out.
4. ⏳ originSessionId back-links — left as a small follow-up.

Both the dashboard search box and the MCP tool benefit since they share the module.