Two tools with reasonable naming (browse, extract) but significant gaps in schema completeness and parameter documentation. Both tools have descriptions (good), but parameter descriptions are minimal or missing. Input schemas are partially visible but lack full type definitions for all parameters. No output schema documentation. Error handling is absent from visible code. The tools accept natural language input which is user-friendly but lacks validation constraints. Session management via session_id parameter is present but undocumented.

• Document output schema for both tools. Specify what fields browse() returns (e.g., success: bool, result: string, steps_taken: int, final_url: string) and what extract() returns (e.g., data: object, extraction_confidence: float, raw_html: string). This enables LLMs to chain results.
• Add parameter descriptions: 'max_steps: Maximum browser interactions (clicks, scrolls, text input). Capped at 50. Default 25 balances thoroughness vs timeout risk.' and 'session_id: Reuse an existing browser session by ID to maintain cookies/state. Omit to create a new session.'
• Add input validation and error messages. Validate task/data_description length (max 4000 chars visible in code), reject empty strings, and return actionable errors: 'Task too vague. Include specific actions: "Click the search box, type "AI news", press Enter, extract top 3 results."'
• Document idempotency: 'browse() is NOT idempotent, retrying the same task may execute duplicate actions. Use session_id to resume interrupted tasks instead of retrying from scratch.'
• Add examples to descriptions: 'browse(task="Search for 'AI news' on HackerNews and return top 3 results", url="https://news.ycombinator.com", max_steps=10)' to guide LLM invocation patterns.

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

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