kimi-code-worker hangs on Shell tool via WebSocket API

Bug Report: kimi-code-worker hangs on Shell tool via WebSocket API

Repository: MoonshotAI/kimi-cli
Version: kimi-cli 1.44.0
Environment: Linux, Python 3.12/3.13, kimi web --host 127.0.0.1 --port 5494

Summary

When using the WebSocket API (kimi web), kimi-code-worker hangs indefinitely after receiving a Shell tool call. The worker remains in busy state with 0% CPU and never returns a ToolResult, causing the session to stall forever.

Steps to Reproduce

1. Start kimi web:

   kimi web --host 127.0.0.1 --port 5494
   

2. Create a session via REST:

   curl -X POST http://127.0.0.1:5494/api/sessions/ \
     -H "Content-Type: application/json" \
     -d '{"work_dir": "/path/to/project", "create_dir": false}'
   

3. Connect to the session via WebSocket and send a prompt that triggers a Shell tool call:

   "what is my IP address?"
   

   (or in Russian: "скажи ip своего хоста")

4. Observe the WebSocket event stream.

Expected Behavior

• ToolCall is received.
• Worker executes the Shell command.
• ToolResult is returned with command output.
• TurnEnd is sent.
• Session completes successfully.

Actual Behavior

• ToolCall (or ToolCallPart streaming fragments) is received.
• StatusUpdate may follow.
• No ToolResult is ever returned.
• Worker process remains in S (sleeping) state with 0% CPU.
• Session status stays busy indefinitely.
• kimi.log shows no Tool Shell completed entry for this session.

Observations

1. Commands themselves are not the issue:

   • hostname -I and curl -s --max-time 5 ifconfig.me execute instantly when run manually in the same shell.

2. Shell tool works in direct CLI mode:

   • When running kimi directly in the terminal (not via kimi web), the Shell tool executes correctly for commands like ruff check, pytest, systemctl, etc.

3. Streaming tool calls:

   • In the WebSocket mode, we observed ToolCallPart events (streaming fragments of a tool call) being delivered before the full ToolCall. The worker may not handle streaming tool calls correctly.

4. Parallel tool calls:

   • In another reproduction, the LLM returned two consecutive ToolCall events (parallel calls). The worker also hung in that scenario.

5. Process state:

   • kimi-code-worker PID shows Sl state, 0% CPU, no child processes.
   • strace shows only epoll_wait loops — the worker is idle, not blocked on I/O.

Wire Protocol Evidence

Hung session (WebSocket):

{"type": "ToolCall", "payload": {"type": "function", "id": "tool_xxx", "function": {"name": "Shell", "arguments": "{\"command\": \"hostname -I\"}"}}}
{"type": "ToolCall", "payload": {"type": "function", "id": "tool_yyy", "function": {"name": "Shell", "arguments": "{\"command\": \"curl -s ifconfig.me\"}"}}}
// ... silence — no ToolResult, no TurnEnd

Working session (direct CLI):

{"type": "ToolCall", "payload": {"type": "function", "id": "tool_zzz", "function": {"name": "Shell", "arguments": "{\"command\": \"ruff check .\"}"}}}
{"type": "ToolResult", "payload": {"tool_call_id": "tool_zzz", "return_value": {"is_error": false, "output": "All checks passed!\n"}}}

Hypothesis

The kimi-code-worker has a bug in its WebSocket event loop when handling:

• Streaming tool calls (ToolCallPart), or
• Parallel/batched ToolCall events

The worker receives the tool call(s) but fails to dispatch them to the tool execution layer, leaving the session stuck in busy state.

Impact

Any WebSocket client (including third-party integrations like Telegram bots) that relies on kimi web will experience permanent hangs whenever the LLM decides to invoke the Shell tool for certain prompts.

Suggested Fix / Workaround

• Ensure ToolCallPart streaming fragments are correctly buffered and assembled into a complete ToolCall before execution.
• Verify that parallel ToolCall events are properly queued or executed without deadlock.
• Consider adding a watchdog/timeout inside kimi-code-worker to abort stuck tool executions.

---

Reported by a downstream integrator using kimi web as a backend for a Telegram bot gateway.

Start by reproducing the hang with kimi web and the WebSocket flow described, then trace kimi-code-worker around ToolCallPart buffering and parallel ToolCall dispatch. Compare the WebSocket event stream with the working direct CLI behavior. Done means Shell calls produce ToolResult events and the session sends TurnEnd instead of remaining busy.