A production-grade CLI Code Agent with ReAct loop, tool orchestration, multi-LLM support, and sub-agent management
Critical failures across all 16 tools. No parameter descriptions exist. Tools like 'bash' and 'write_file' expose destructive operations with zero input validation guidance. spawn_agent/spawn_agent_async have slightly better descriptions but still lack parameter detail. No output schemas documented. Error handling is generic ('isError: true') with no recovery guidance. This server would not pass code review.
Execute bash commands
Close a sub-agent and release its resources. After closing, the agent_id is no longer valid. Always close sub-agents when done to prevent resource leaks. Parameters: agent_id: The sub-agent ID to close.
Create a reusable skill/function
Edit file contents with line-based operations
Retrieve memory/notes from the agent's memory store
List directory contents
Read file contents
Store a note in the agent's memory
ALL 16 tools have empty input schemas: properties: {}, required: []. This prevents LLMs from understanding what parameters each tool accepts.
14 of 16 tools have descriptions under 50 characters (e.g., 'Read file contents', 'Write or create a file'). These are too terse to guide LLM tool selection.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 38 | 2026-07-28+ | v2 |
Search for files matching patterns
Search through conversation history
Send a message to a sub-agent to start or continue its task. This is NON-BLOCKING: the sub-agent begins executing but this tool returns immediately. After calling send_input, you MUST call wait_agent to get the result. Parameters: agent_id: The ID returned by spawn_agent_async. message: The task or follow-up message for the sub-agent.
Create and execute a sub-agent synchronously (blocking)
Create a new sub-agent and return its ID immediately (non-blocking). The sub-agent starts in idle state. Use send_input to give it a task. Use this when you want to run multiple sub-agents in parallel, or when you need to interact with the sub-agent across multiple turns. Returns: agent_id (string) — use this in send_input / wait_agent / close_agent.
Update the current task plan
Wait for a sub-agent to complete its current task and return the result. This BLOCKS until the sub-agent finishes. Call this after send_input to retrieve the result. Parameters: agent_id: The sub-agent ID to wait for.
Write or create a file
Destructive tools (bash, write_file, edit_file, delete operations) lack error handling guidance and confirmation patterns. No dry-run, no recovery instructions. An LLM could execute 'rm -rf /' with no safeguards.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., agent_id from spawn_agent_async). Responses are unstructured.
Error responses are generic (isError: true) with no recovery guidance. Example: 'agent_id is required' does not explain how to obtain one or retry.