Sanzaru provides 22 tools across video, audio, and image domains with explicit schema definitions visible in server.py. Tool names follow verb-noun conventions (create_video, list_audio_files, transcribe_audio) and most descriptions are present and non-trivial (avg ~120 chars). However, significant gaps emerge: (1) Many parameter descriptions are sparse or missing detail about constraints (e.g., 'seconds' in create_video lacks range info; 'offset' and 'chunk_size' in _get_media_data have minimal guidance). (2) Output schemas are NOT documented anywhere in the visible source, no specification of what fields are returned or their structure. (3) Several tools mix concerns or have unclear responsibilities (e.g., create_image vs generate_image both do image generation; _get_media_data is marked 'do not call directly' but remains exposed). (4) Error handling is minimal, no evidence of recovery guidance or actionable error messages. (5) Tool definitions reference external services (OpenAI Sora, Whisper, DALL-E, ElevenLabs) but do not document failure modes, timeouts, or prerequisites. The server demonstrates competent naming and basic schema presence, placing it solidly in the C+ range rather than B-.

• Document output schemas for all tools. For each tool, add a 'Returns' section specifying field names, types, and meaning. E.g., create_video should document that it returns {video_id: string, status: string, created_at: ISO8601}.
• Remove _get_media_data from public tool registration. It is a private implementation detail. If chunked media fetching is needed, wrap it in a public tool with clear naming.
• Consolidate create_image and generate_image into a single canonical image generation tool. Document when to use reference_image vs style, and make clear which are optional.
• Add parameter bounds and formats to tool descriptions. E.g., 'seconds (required, 5-120): video duration'; 'limit (default 20, max 100): results per page'.
• Add error recovery guidance to tools calling external APIs. Document: typical failure modes (rate limit, timeout, invalid input), suggested retries, and fallback strategies. E.g., 'If create_video fails with rate_limit_exceeded, wait 30 seconds and retry.'
• Expand parameter descriptions to include constraints and examples of valid input. E.g., 'filename (string): Media filename in storage (e.g., video.mp4, song.mp3). Supports ASCII alphanumeric, dash, underscore, dot. Max 255 chars. No path traversal (..) allowed.'
• Add permission and audit notes to destructive tools. E.g., delete_video: 'Permanently removes video from OpenAI storage. This action cannot be undone. Requires admin permission. All deletions are logged.'
• Document API prerequisites and dependencies. E.g., create_video requires OPENAI_API_KEY to be set. transcribe_audio requires Whisper model availability. Mention fallback providers where applicable (TTS can use ElevenLabs).

Inferred effective spec: 2026-07-28+.

• Current-spec patterns present: stateless request handling, structured tool output, Streamable HTTP