This server has serious quality gaps across naming, descriptions, and schema completeness. While all four tools are properly registered with FastMCP and have basic docstrings, the implementation exhibits multiple critical issues: (1) tool names lack action verbs in three of four cases (read_existing_rules is acceptable, but generate_project_rules, save_rules_file, update_existing_rules are weak); (2) descriptions are present but generic and underpowered, they do not explain WHEN to use each tool vs. alternatives, lack dependency hints, or guide multi-step composition; (3) parameter schemas are incomplete, most parameters lack proper JSON Schema type definitions beyond string/array, missing min/max constraints for numeric/array types, and no enums where appropriate; (4) output schemas are not documented, callers cannot predict the structure of responses and must guess or reverse-engineer; (5) error handling is minimal, exceptions are caught but responses are unstructured and do not guide LLM recovery; (6) critical composition issues, the tools operate in sequence but responses do not carry forward all necessary identifiers for chaining calls. The STDIO transport compounds these issues by making the server untestable in production scenarios.
根据项目类型和功能特性生成新的项目规则文件
读取现有的规则文件内容
保存规则文件到指定目录
更新现有的规则文件内容
Weak tool naming: 'generate_project_rules', 'save_rules_file', and 'update_existing_rules' do not start with clear action verbs. 'generate_project_rules' could mean compile, create, or synthesize. 'save_rules_file' is passive. Recommend: 'create_project_rules', 'write_rules_file', 'patch_rules_file'.
Missing output schema documentation. None of the four tools document their return types. read_existing_rules returns a dict with 'status', 'data', 'message' keys, but the LLM must infer structure. generate_project_rules returns a string; save_rules_file returns a string; update_existing_rules has no explicit return type shown. Without documented schemas, LLMs cannot plan downstream calls or extract relevant fields.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 43 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 40 | - | v1 |
Input parameter schemas lack proper type definitions and constraints. (1) 'rules_path' is string with no min/max length, no path validation pattern, no enum of valid paths. (2) 'features' is array but items lack additional constraints; no min/max items, no enum of valid feature names. (3) 'language' defaults to '中文' but no enum constraining values to valid options. (4) 'content' in save_rules_file has no length constraint. (5) 'updates' object in update_existing_rules is untyped, the nested 'properties' are documented in description but not in JSON Schema additionalProperties or patternProperties. These omissions invite invalid inputs and force LLMs to guess.
Descriptions are generic and do not guide tool selection. 'read_existing_rules' says '读取现有的规则文件内容' (read existing rule file content), any LLM could call this or another discovery tool. Descriptions should explain: When is this preferred over list_rules? What does it return that others don't? Dependency hints are absent (e.g., 'Call this before generate_project_rules to check what rules already exist').
Error handling is minimal and unstructured. read_existing_rules returns {'status': 'error', 'message': '...'} on failure, but this is inconsistent with the success case {'status': 'success', 'data': {...}}. Errors do not guide recovery: when 'rules_path' does not exist, the error should suggest 'Use save_rules_file() to create the directory first.' generate_project_rules returns a raw string on error with no structure, LLMs cannot parse or categorize the failure. No timeouts, no retryable vs. fatal classification.
Composition chain is broken. read_existing_rules returns file list with 'name', 'path', 'size', 'content_preview' but update_existing_rules requires 'file_path' (full path string). The response does include 'path' relative to rules_dir, but no absolute path or content, forcing the LLM to manually construct file_path or call read twice. If read_existing_rules returned {'file_path': '...', 'rules_dir': '...'} or absolute path, the chain would work directly.
Parameter descriptions lack specificity. 'features' is described as 'Project feature list (e.g. ["authentication", "database", "api"])' but the examples are presented as the only valid options, when in reality any feature name is accepted. Recommend: 'Array of feature names (e.g., authentication, database, api, caching). Any string is accepted; common names trigger specialized rules.' This prevents LLMs from limiting themselves to the listed examples.