Soroka exposes 9 tools for knowledge base search and retrieval. All tools are read-only except delete_note (destructive). Tool names follow verb_noun convention consistently (search, get_by_id, list_recent, get_attachment, delete_note, find_similar, get_context, get_by_ids, stats). Descriptions are present and moderately detailed (average ~80 chars). However, several parameters lack descriptions, and output schemas are not documented in the source. The delete_note tool references audit logging and soft-delete semantics but provides no recovery mechanism or confirmation pattern. Pagination support is present in search (limit, offset) and list_recent (limit), though not all list operations document per-item success/failure handling.

• Document output schema for each tool. Example: search() returns [{note_id: int, content: string, kind: enum, created_at: ISO8601, relevance_score: float}, ...]. Specify which fields are always present vs. optional.
• Add dry-run parameter to delete_note: delete_note(note_id, reason, dry_run=false). Return what would be deleted without committing. Enables LLM to preview before executing.
• Define bounds on numeric parameters: get_context(window: 1-50 default 3), search(limit: 1-100 default 5), list_recent(limit: 1-100 default 20). Prevents runaway agent requests.
• Add error response documentation: 'If note_id does not exist, returns {error: "note_not_found", available_ids: [1, 2, 3]}' so LLMs can offer alternatives instead of dead-ending.
• Clarify stats() output: does it return {total_notes: int, notes_by_kind: {text: 10, web: 5, ...}, date_range: {earliest: ISO8601, latest: ISO8601}, activity_by_period: {...}}? Document structure explicitly.
• Add parameter description for get_attachment explaining base64 encoding and encoding error handling.
• Consider splitting search into search_by_text (free-form query) and search_by_metadata (filter by kind, date range) if they have different use patterns; current search combines both, which is good, but the combined flexibility should be highlighted in the description.