自进化 JS 逆向工作流 MCP Server — 集成 MemOS / GitNexus / Stagehand / js-reverse-mcp
Four tools with complete schemas and descriptions, but significant quality gaps. Tool names lack clear action verbs (evolve_js, query_knowledge, inject_knowledge are vague). Descriptions are present but generic, they describe WHAT without explaining WHEN to use or dependencies. Parameters have types and descriptions, but lack validation constraints (e.g., successThreshold 0-1 not enforced in schema). Output schemas are undocumented, callers don't know what fields to expect. Error handling is absent from tool definitions. Security concerns: inject_knowledge accepts arbitrary JavaScript code as 'fix' parameter without sanitization or sandboxing warnings. No pagination support despite knowledge base queries that could return large result sets. Tool composition is weak, no clear data flow between tools (e.g., evolve_js doesn't return sessionId for reuse in subsequent calls).
对目标 URL 启动自动进化逆向工作流,直到反混淆代码匹配率达到阈值(默认 100%)。内置自学习:每次调用都会积累经验,越用越快。
查看 GitNexus 沙箱仓库中所有进化分支的状态(开放中 / 已合并 / 已放弃)。
手动向知识库注入人工经验(混淆模式或自定义策略),干预进化方向,加速收敛。
查询系统积累的知识库,包括混淆模式、成功历史、失败路径、策略候选及整体统计。
Tool names lack action verbs. 'evolve_js' is vague (evolve how? to what end?). 'query_knowledge' and 'inject_knowledge' are generic. LLMs cannot infer intent from names alone.
Output schemas are completely undocumented. evolve_js returns JSON with 'success', 'accuracy', 'iterations', 'branchName', 'knowledgeGained', 'codePreview', but this is inferred from code, not declared in tool definition. LLMs cannot plan downstream calls without knowing return structure.
inject_knowledge accepts arbitrary JavaScript code in 'fix' parameter with no sanitization, sandboxing, or security warning. This is a code injection vector. Description must warn of risks and implementation must validate/sandbox execution.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | <=2025-11-25 | v2 |
No pagination support. query_knowledge returns up to 'limit' items (default 20) but no indication of total count, next_cursor, or whether results are truncated. Large knowledge bases will be silently capped.
Descriptions lack WHEN/WHY context. 'evolve_js' says WHAT it does but not when to call it vs. other tools, what prerequisites exist, or what 'self-learn' means operationally. Descriptions should be 50-200 chars and prompt-engineered for LLM selection.