The server provides two read-only documentation tools with solid naming (verb_noun pattern) and comprehensive descriptions. However, parameter documentation is inconsistent, output schemas are not formally documented, and error handling lacks recovery guidance. The tools follow good composition principles (single responsibility) but fall short of production-grade polish.

• Document the output schema for oci_search_documentation in the docstring: 'Returns JSON with structure: {"pagination": {"page": int, "limit": int, "total": int}, "results": [{"url": string, "title": string, "description": string, "rank": float}]}'. This tells the LLM what fields to extract.
• Document the output schema for oci_read_documentation: 'Returns JSON: {"content": string (markdown), "truncated": bool, "total_lines": int, "next_start_index": int}'. This clarifies when pagination is needed and what start_index to use next.
• Add recovery guidance to error handling: catch specific exceptions (network timeout, invalid URL, rate limit) and return actionable messages like 'URL not found (HTTP 404). Verify the documentation exists by searching for it with oci_search_documentation().' instead of generic fallback.
• Clarify 'start_index' units in oci_read_documentation parameter description: 'Line number (0-indexed, where first line = 0) to start from. Use the next_start_index from a previous truncated response to continue reading.'
• Document pagination for oci_search_documentation: 'Pages are 1-indexed. Results for page=2, limit=3 are items 4-6 (second batch). When results.length < limit, you have reached the end.'
• Replace 'max_lines' constraint 'lt=1000000' with 'le=999999' and update description to: 'Maximum number of lines to return (1 - 999,999). For very long documents, make multiple calls with different start_index values.'
• Add a 'Best Practices' section to the tool descriptions reminding LLMs: 'For search results, always verify the returned URLs match the requested topic before reading. For long documents, stop reading early if you've found the answer, each line costs tokens.'

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

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry