ViewGraph provides 41 tools with consistent naming (verb-prefixed) and parameter schemas. However, 39 of 41 tools are READ_ONLY utilities (auditing, analysis, comparison), leaving the core domain incomplete. Tool descriptions are adequate (all present, ~30 - 100 chars each) but lack depth, most do not explain WHEN to call them, dependencies, or expected outputs. Parameter descriptions are minimal ('Capture filename', 'URL to capture') without format constraints, ranges, or examples of valid input. No input validation guidance visible in tool definitions. Output schemas are not documented, we can infer they return structured data, but no formal schema definitions are visible in the source snippets. Tool composition is strong (names follow verb_noun convention, clear distinctions), but error handling is not addressed in the code samples. Two write operations (request_capture, resolve_annotation, set_baseline) lack confirmation or dry-run patterns despite modifying state. Security checks (permission gates, secret injection) are not visible. Overall, this is a solid mid-tier implementation with good naming discipline and parameter coverage, but insufficient description depth and missing output schema documentation prevent a higher score.

• Expand tool descriptions to 80 - 120 characters, including: (1) What does it do? (2) When should the LLM call it vs similar tools? (3) What does it return? Example: 'get_capture: Retrieve the full DOM, screenshot, and metadata for a saved capture. Call this to inspect element details, run accessibility audits, or compare to baselines. Returns capture object with dom, screenshot, metadata, and annotations.'
• Document output schemas explicitly. For each tool, specify: response type (object/array), key fields (with types and descriptions), and pagination structure if applicable. Example: 'Returns {captures: [{filename: string, url: string, timestamp: ISO8601, elementCount: number}], total: number, nextCursor?: string}'
• Add format/constraint descriptions to parameters. E.g. 'urlFilter: substring to match in capture URLs (case-insensitive, e.g. "github.com")'; 'limit: number of results to return (1 - 100, default 20)'; 'role: ARIA role name (e.g. "button", "link", "form"), see https://w3c.github.io/aria/'
• Implement confirmation pattern for write operations. Add optional 'dryRun: boolean' or 'confirm: boolean' to request_capture, resolve_annotation, and set_baseline. For destructive operations, return a confirmation token the agent must pass back in a confirm_* call.
• Add error handling documentation to tool descriptions. E.g. 'Returns error if capture file not found, suggest list_captures() to see available files; returns error if no permission to access capture, contact administrator.'
• Add scope declarations to write operations. E.g. 'request_capture requires scope: write:captures; resolve_annotation requires scope: write:annotations.' Document these in server initialization so clients know what permissions to request.

Inferred effective spec: <=2025-11-25.

• Relies on Sampling (deprecated) - integrate directly with the LLM provider API