16 tools with complete input schemas and descriptions. Naming is action-verb based (ask_, gemini_, generate_, web_) and mostly clear. Descriptions average ~150 chars, meeting baseline (194 char target). However, parameter descriptions are sparse or missing in several tools (e.g., gemini_summarize lacks 'content' description; gemini_count_tokens has minimal context). Output schemas are undocumented, LLMs cannot predict return structure. Error handling is absent from visible code; no recovery guidance. Security: tools accept file paths without validation, risking path traversal. Composition is reasonable (16 focused tools), but some overlap (ask_gemini vs web_ask) and missing chaining IDs in responses.

• Document output schemas for all tools. Example: ask_gemini returns {response: string, sources?: [{url: string, title: string}], model_used: string}. Include pagination fields (total_count, next_cursor) for list-like responses.
• Add descriptions to all parameters. For gemini_summarize, describe 'content' as 'The text to summarize (supports plain text, markdown, code)'. For gemini_count_tokens, clarify 'text' as 'Source text (code, prose, or mixed; ~4 chars per token).'
• Implement input validation and error handling. Validate file paths (reject absolute paths outside project root), timeout values (1 - 3600 seconds), and model names (enum: flash-lite, flash, pro). Return structured errors: {error: string, code: string, retryable: boolean, suggestion: string}.
• Consolidate ask_gemini and web_ask. Remove web_ask or clearly document it as 'Fallback for ask_gemini when CLI unavailable, prefer ask_gemini.' Reduce LLM decision overhead.
• Add tool annotations (readOnlyHint, destructiveHint) to distinguish safe tools (ask_gemini, gemini_search) from write tools (generate_image, generate_video). Helps agents plan safely.

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

• Relies on Stateful initialize / Mcp-Session-Id (removed; protocol is stateless) - make each request self-contained