基于 ReAct 循环的代码开发助手 - An AI-powered code development assistant based on ReAct loop with MCP tool integration
IntelliAgent has 16 tools with mixed quality. Naming is generally verb-based (read_file, write_file, run_shell), but descriptions vary widely in completeness. Most tools lack detailed parameter descriptions and output schemas. File tools have reasonable descriptions but lack explicit output schema documentation. Task and agent team tools have minimal descriptions (under 50 chars). Error handling is present in the executor but not documented in tool definitions. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Parameters often lack type constraints and validation guidance.
创建新的 Agent。
删除指定的 Agent。
异步编辑文件内容,精确替换指定字符串。支持单次替换和全部替换两种模式,替换后自动写回文件。
获取指定 Agent 的详细信息。
获取当前 Agent 的联系人列表。
获取当前时间(中国标准时间 CST)
加载指定 skill 的完整指令。
异步读取文件内容。支持工作区边界检查、文件大小截断,优先使用 aiofiles 异步读取。
Task and agent team tools have minimal descriptions (under 50 chars). 'task_write' is '批量创建任务' (8 chars), 'send_message' is '向另一个 Agent 发送消息' (12 chars). LLMs cannot determine when to use these tools or what they return.
No output schemas documented for any tool. Callers cannot know what fields to expect (e.g., does task_add return task_id, title, status?). This forces LLMs to guess and breaks tool chaining.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 46 | 2026-07-28+ | v2 |
接收发送给当前 Agent 的消息。
异步执行 shell 命令并返回结果。使用 bash -lc 执行命令(如果 bash 可用),支持超时控制。命令执行超时后会强制终止进程。
向另一个 Agent 发送消息。
创建单条任务。
标记任务为已完成。
更新单条任务,只更新传入的非空字段。
批量创建任务。
异步写入文件内容。自动创建父目录,支持工作区边界检查,优先使用 aiofiles 异步写入。
No tool annotations despite clear risk levels. run_shell is IRREVERSIBLE, delete_agent is DESTRUCTIVE, write_file is WRITE, but no readOnlyHint/destructiveHint/idempotentHint in schema. LLMs cannot reason about safety.
Parameter descriptions are sparse or missing. 'path_policy' in read_file/write_file/edit_file is described as '路径边界策略,优先于 workspace_root' but no type, format, or example. 'tasks' in task_write is '任务列表' with no schema for the JSON array structure.
Error handling is implemented in ToolExecutor but not documented in tool definitions. No guidance on retryability, user-fixable errors, or recovery steps. LLMs receive raw error JSON without context.