Expose DeepSeek Harness agent capabilities as an MCP server, so an external MCP client (e.g. Hermes) can drive Harness to execute coding tasks. Hermes = brain, Harness = arms.
Server exposes 7 tools with basic schemas and Chinese descriptions. All tools have input schemas with type definitions, but descriptions are minimal (10-50 chars) and lack LLM-optimized guidance on WHEN to use each tool or WHAT happens next. No output schemas documented. Error handling is absent, tools return unstructured text via generic `out()` wrapper. Tool naming mixes verbs (agent_run, task_inbox) with nouns (harness_list_tools, task_result), reducing clarity. No parameter validation, no recovery guidance, no idempotent hints. Composition is reasonable (task_inbox → task_result chain), but descriptions don't explain this flow.
同步执行任务(改代码/分析/跑命令), 返回结构化结果
把会话归组到其 cwd 对应的工作区(手动补给站)
验证 MCP server 连通
列出 Harness 工具注册表
给已有会话改名
Hermes push 结构化任务(任务+记忆上下文)到 Harness 队列, 异步执行, 返回 taskId
取回任务的结构化结果(changes/verification/leftovers)
Descriptions are too short (10-50 chars) and lack LLM-optimized guidance. Chinese descriptions like '验证 MCP server 连通' (verify MCP server connectivity) do not explain WHEN to call the tool, WHAT it returns, or dependencies. Baseline for A+ tools is 50-200 chars with explicit context.
No output schemas documented. Tools return generic text via `out({content: [{type: 'text', text: content}]})`. LLMs cannot infer what fields to expect (e.g., does task_result return changes, verification, leftovers as separate fields or unstructured text?). Baseline: 100% of A+ tools document return types.
No error handling or recovery guidance. Tools throw generic errors (e.g., 'session not found for resume') without actionable next steps. LLMs cannot determine if an error is retryable, user-fixable, or fatal. Baseline: error responses must tell the LLM what to do next.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | <=2025-11-25 | v2 |
Tool naming inconsistency. 'harness_list_tools' uses noun-first pattern (list_tools) while 'agent_run' uses verb_noun. 'task_result' is ambiguous, does it retrieve, update, or delete? Baseline: 90% of A+ tools start with action verbs (get_, create_, update_, delete_, search_, list_).
Parameter descriptions are minimal or absent. 'context' param in agent_run and task_inbox lacks detail on format, size limits, or expected structure. 'cwd' is not validated for path traversal. Baseline: every parameter needs a description explaining what it controls and constraints.