This server implements 4 tools for Bilibili subtitle extraction and video search. Naming is generally clear and verb-based (fetch_subtitles, search_videos, get_video_info, fetch_subtitles_with_asr). All tools have descriptions present and reasonable length (50-150 chars). However, schema quality is mixed: while input parameters are defined with types and descriptions, output schemas are NOT documented in the visible code, LLMs cannot plan downstream calls or extract structured data reliably. Error handling exists at the code level but is not visible in tool definitions. The fetch_subtitles_with_asr tool combines multiple concerns (fetch + fallback to ASR) which could be decomposed. Parameter descriptions are present but some lack detail on constraints (e.g., asr_model_size lacks validation guidance). CRITICAL: the fetch_subtitles and fetch_subtitles_with_asr tools both accept url OR bvid parameters, creating mutual exclusivity that is handled in code but not declared in schemas, LLMs will pass both. No tool annotations (readonly, destructive hints) are present despite fastmcp support.

• Add explicit output schema documentation for all four tools. Specify fields, types, and meanings. Example for fetch_subtitles: {subtitle_text: string, language: string, bvid: string, cid: integer}.
• Convert fetch_subtitles_with_asr into two composable tools: fetch_subtitles (returns subtitle_url + language) and transcribe_video_audio (accepts bvid, returns transcribed_text + timestamps). This follows pattern:tool (one action per tool) and allows agents to decide when to use fallback transcription.
• Add enum constraint to asr_model_size in schema. In fastmcp, use @mcp.tool(input_schema={'asr_model_size': {'type': 'string', 'enum': ['tiny', 'base', 'small', 'medium', 'large']}}).
• Declare mutual exclusivity in parameter descriptions: 'Provide exactly one of url or bvid. If both are provided, an error will be returned.' Ideally, refactor to accept only one parameter type (bvid) and resolve short URLs server-side before tool invocation.
• Add tool annotations to all tools. In fastmcp: use @mcp.tool(..., read_only_hint=True) on all four tools to signal they are safe to call without user confirmation.
• Document search_videos pagination. Add constraints: 'page must be >= 1 and typically <= 100. Returns paginated results with total_count. Use page+limit to fetch subsequent batches.'
• Expand preferred_lang documentation with examples: 'ISO 639-1 language code (e.g., zh-CN, zh-TW, en, ja). Defaults to zh-CN or BILIBILI_PREFERRED_LANG env var if set. Falls back to first available language if requested language not found.'

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

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