session-recall

# Session Recall Use the shared Claude Code + Codex index so the user does not have to re-explain prior work. Treat each result's `source` (`claude` or `codex`) as provenance, not relevance. ## Recall workflow 1. Start with `recall_search`; use `recent_sessions` first when the request is about the latest state or index freshness. 2. When the request names one day, pass `on_date`; for a range, pass inclusive `start_date` / `end_date`. The server defaults to the user's computer timezone, so pass `timezone` only when the user requests an override. Preserve the filter across `recent_sessions`, `recall_search`, and `grep`; do not rely on putting the date into the semantic query. 3. Pass `scope_cwd` for repo-local questions. Omit it for cross-project recall; retry globally if a scoped search is thin. 4. Treat Claude Code and Codex as one history by default. Omit `source` so both are searched; results retain `source="claude"|"codex"` as provenance. Filter by source only when the user explicitly names a host or provenance is material, and preserve that filter while drilling in. 5. For deeper grounding, inspect the best anchors with `expand_around`, walk with `step`, and use `grep` for exact identifiers that semantic search misses. 6. `recall_search` answers with `{"anchors": [...], "degraded": null | str}`. When `degraded` is set, the embedding provider was unreachable and only literal word matching ran: say so to the user, retry with the exact identi...