Single tool 'run_faceswap' has a well-structured input schema with 10 parameters, all typed and most described. Tool description is adequate (95 chars) but generic, lacks WHEN to use it, prerequisites, or state-change clarity. Parameter descriptions vary in quality: 'sourceImagePath' and 'targetMediaPath' are clear, but 'executionProviders' lacks guidance on when to use which provider. Output schema is not documented, LLMs cannot predict what fields the response contains. No error handling guidance provided. Naming is verb-forward ('run_faceswap') but could be more specific (e.g., 'swap_faces_in_media'). Missing idempotency and retry semantics for a long-running operation.
Perform face swap on images and videos. Supports GIFs and MP4 files.
Output schema not documented. LLMs cannot predict response structure (success/failure, output file path, metadata). Breaks downstream tool chaining and forces agents to guess field names.
Tool description lacks state-change clarity and prerequisites. Does not state 'This operation modifies files on disk' or 'Requires FaceFusion models to be installed'. Agents cannot assess side effects or prerequisites.
Parameter 'executionProviders' has enum values but no guidance on when to use coreml vs cpu vs cuda. LLMs will guess or pick arbitrarily, risking performance or compatibility issues.
No error handling guidance. If face detection fails, file is corrupted, or output path is invalid, LLMs receive no recovery hints. Errors should guide next steps (e.g., 'Try a different source image' or 'Check output directory permissions').
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | 2026-07-28+ | v2 |
Long-running operation lacks timeout, progress, or cancellation semantics. No documentation of expected duration or how to handle hung requests. Agents may block indefinitely.