Three tools with complete schemas and descriptions, but descriptions are terse (avg 95 chars vs baseline 194). Tool names follow verb_noun pattern (dig, dig_read, dig_state) but lack clarity on distinctions. All tools are READ_ONLY with no destructive operations, reducing risk. Schemas are well-formed with proper types and required fields. Parameter descriptions are present but minimal. No error handling guidance or recovery paths documented. Output schemas not explicitly documented.

• Expand tool descriptions to 150-250 chars. For dig_state, add: 'Shows working-state snapshots (recent messages, todos, touched files) at each compaction point. Use this to understand context and decisions at a specific moment in the session history.'
• Document output schemas in code comments or a schema registry. Example for dig(): 'Returns array of {match_id, file, offset, role, timestamp, snippet}'. This enables LLMs to chain dig() → dig_read() correctly.
• Add error handling examples: 'If no matches found, returns empty array. If match_id is invalid, returns error with available match_ids from the session.'
• Add bounds to numeric parameters: 'limit: integer, 1-50 (default 5). Maximum matches to return.'
• Rename dig_read to 'read_archived_message' and dig_state to 'list_state_snapshots' for clarity and to follow verb_noun convention more strictly.