Startline AI has 8 well-named, read-only tools with clear descriptions (avg 120 chars, within baseline 194). All tools have input schemas with type definitions and parameter descriptions. However, output schemas are not documented, responses are inferred from code inspection. Error handling is minimal (only search() returns error objects; others return dicts with 'error' key inconsistently). No tool annotations (readOnlyHint, idempotentHint). Parameter descriptions are good but lack format constraints (e.g., event codes like 'AUS-GP-2026' are documented by example, not enum). The vault.find() fuzzy matching and synonym expansion show thoughtful design, but composition could be tighter, read_doc and search both require event disambiguation.

• Add explicit output schema documentation to every tool docstring. E.g., search() should document: 'Returns: list[dict] with keys: doc (str), where (str), note (str), event (str), snippet (str), status (str, optional).'
• Replace example-based event code documentation with enum constraint or regex pattern in schema. E.g., event parameter: type=string, pattern='^[A-Z]{3}-[A-Z]{2}-\d{4}$', description='Event code, e.g. AUS-GP-2026. See list_events() for valid codes.'
• Enhance error messages with recovery guidance. E.g., event_overview() on missing event: 'Event AUS-GP-2026 not found. Valid events: AUS-GP-2026, MON-GP-2026. Call list_events() to see all.'
• Add readOnlyHint: true to all 8 tool definitions in FastMCP registration to signal safety to agents.
• Constrain read_doc() pages parameter with pattern: '^\d+(?:\s*-\s*\d+)?$' in schema, and update description to: 'Page range: single page (e.g. "3") or range (e.g. "3-5"). Required for documents >60KB.'