JSON parsing error when tool call arguments contain control characters

[AlejandroEsquivel]

What version of Kimi Code CLI is running?

1.17.0

Which open platform/subscription were you using?

Moonshot AI Open Platform

Which model were you using?

kimi-2.5

What platform is your computer?

Linux 6.1.158 x86_64 unknown

What issue are you seeing?

When the LLM generates tool_call.function.arguments containing literal control characters (e.g. unescaped newlines in multi-line git commit messages), json.loads() raises a JSONDecodeError because Python's JSON parser rejects unescaped control characters (U+0000 through U+001F) by default.

This causes a cascade of failures:

1. Tool execution fails — json.loads() in toolset.py:123 and simple.py:122 raises JSONDecodeError
2. Session history is poisoned — the malformed arguments string is persisted to the context file
3. Session becomes irrecoverable — on restore, json.loads(line) in context.py:40 fails, and on every subsequent API call _convert_message() sends the malformed arguments back to the API, which returns HTTP 400

What steps can reproduce the bug?

1. Start a kimi-cli session
2. Ask the LLM to run a shell command that includes a multi-line string, e.g.:

   Run: git commit -m "feat: add feature
   
   This is a multi-line
   commit message"
   

3. The LLM generates a Shell tool call with literal newlines in the JSON arguments string
4. Observe the error:

   ERROR: Error parsing JSON arguments: Invalid control character at: line 1 column 264 (char 263)
   

5. The session is now broken — resuming it will fail with the same error or HTTP 400

What is the expected behavior?

No response

Additional information

Example trace

{
  "role": "assistant",
  "content": [],
  "tool_calls": [
    {
      "type": "function",
      "id": "functions.Shell:22",
      "function": {
        "name": "Shell",
        "arguments": "{\"command\": \"git commit -m \\\"feat: add feature\n\nMulti-line message\\\"\"}"
      }
    }
  ]
}

{
  "role": "tool",
  "content": "<system>ERROR: Error parsing JSON arguments: Invalid control character at: line 1 column 264 (char 263)</system>",
  "tool_call_id": "functions.Shell:22"
}

[Br1an67]

I've looked into this and the root cause is that json.loads() uses strict=True by default, which rejects control characters (newlines, tabs) inside JSON string values. When the LLM returns tool call arguments containing these characters, the parse fails and the corrupted line in context.jsonl prevents session restoration.

The fix is straightforward: use json.loads(..., strict=False) in the critical parsing paths — tool call arguments (toolset.py), session context restore/revert (context.py), tool argument display (tools/__init__.py, debug.py).

I have a PR ready if you'd like to take a look.