Static source inference · medium confidence · detected: Logging
Deprecated protocol patterns detected
Summary
This server has critical gaps in definition quality across all tools. Descriptions are present but extremely brief (averaging 15-20 chars in Japanese), far below the 50-200 char baseline for LLM optimization. No input schemas are visible in the provided source code, only parameter names and basic types are shown. Tool naming follows verb_noun patterns (openFile, closeFile, saveFile) but uses camelCase instead of snake_case convention. More critically, there is no evidence of output schema documentation, error recovery guidance, or actionable error messages. The server exposes a custom 'bridge_mode' parameter across all tools, which is an implementation detail that should be abstracted server-side. Security concerns include the executeScript tool accepting arbitrary JavaScript, and no visible input validation or sanitization. The code snippet shows HTTP transport (FastAPI) and basic handler registration, but tool definitions lack the formality required for agentic use.
All tool descriptions are under 25 characters (mostly 10-20 chars in Japanese translation). This violates the 50-200 char baseline and prevents LLM tool selection optimization. Descriptions must include WHAT the tool does, WHEN to use it, and prerequisites.
No input schemas are visible in the source code. The metadata shows parameter names and type hints, but no explicit JSON Schema with 'type', 'properties', 'required' fields, or schema constraints. Cannot verify minimum/maximum bounds on numeric parameters, enum constraints, or format specifications.
Expand all tool descriptions to 50-200 characters. Example: 'openFile' should be 'Open a PSD file in the currently running Photoshop instance. Requires Photoshop to be running. Use this before modifying or exporting layers. Returns success status or error message.'
Document full JSON Schema for inputs: include 'type', 'properties', 'required', 'enum' (for bridge_mode), and numeric bounds. Example: bridge_mode should be enum: ['applescript', 'uxp', 'grpc'] with default specified.
Remove 'bridge_mode' from tool parameters. Move to server configuration or environment variable. Abstract this implementation detail away from the LLM interface.
Add output schema documentation to every tool. For getDocumentInfo, specify returned fields: { name: string, width: integer, height: integer, layers: array, ... }.
Add error recovery guidance to all error responses. Replace 'Photoshop failed to open file' with 'Failed to open file: File not found at /path/to/file. Check file path and ensure Photoshop is running. Retry with correct path.'
Implement input validation with constraint violations in error messages. For executeScript, validate script length, syntax (if possible), and return specific error: 'Script exceeds 10,000 chars. Reduce size and retry.'
Add destructive operation warnings. Tools like saveFile, closeFile (with save_changes=true), and deleteLayer should include 'This operation cannot be undone' in the description and declare destructiveHint in tool annotations.
Output schemas are not documented. Tool descriptions do not state what fields are returned or their types. For example, getDocumentInfo returns 'info' but the structure is undocumented. LLMs cannot plan downstream calls or extract the right data without knowing response schema.
All tools expose a 'bridge_mode' parameter as user-facing input. This is an implementation detail (switching between AppleScript, gRPC, UXP) and should be abstracted server-side via configuration. LLMs should not reason about bridge modes; it creates confusion and allows misconfiguration.
executeScript tool accepts arbitrary JavaScript code as input. No description of sandbox constraints, API limitations, or error boundaries. This is a critical security and reliability risk, an LLM could generate malicious or invalid scripts. Needs explicit security documentation and input validation.
Error handling in code shows generic HTTPException(500, 'Photoshop failed to X') without actionable recovery guidance. Errors do not categorize as retryable, user-fixable, or fatal. LLM has no instruction on next steps (retry, ask user, abort).
Tool naming uses camelCase (openFile, closeFile) instead of snake_case convention (open_file, close_file). While functional, this deviates from the 549-tool baseline where 90% of A+ tools use snake_case verb_noun naming.
Tool 'healthz' has an atypical name for a health check. Conventional names like 'get_server_health' or 'check_system_status' are clearer. 'healthz' is Kubernetes jargon; LLMs may not recognize it as a diagnostic tool.
Parameter descriptions are minimal (10-15 chars). E.g., 'ブリッジモード' (bridge mode) does not explain what modes are available, when to use each, or default behavior. Parameter descriptions must state format, range, and allowed values per the 72-char baseline.
Document the 'save_changes' parameter in closeFile with boolean constraint and default value (e.g., default false to prevent accidental data loss).
For generateThumbnail, specify enum constraints: format = ['jpeg', 'png'], quality = 0-100 integer, width/height = 10-4000 pixels.
Add parameters that accept human-readable identifiers alongside IDs. For example, layer parameter in exportLayer could accept layer name or layer_id to reduce lookup overhead.
Implement pagination for any list/search operations. If submitJob or getJobStatus can return multiple results, add limit and offset/cursor parameters.
Create separate tools for read vs. write operations to support read-only agent configurations. Group getDocumentInfo, generateThumbnail, healthz, getJobStatus as 'read_* tools'; create matching write_* tools with explicit permission warnings.