bug: Tool call arguments double-encoding breaks array/dict parameters (Moonshot API)

Problem

Tool calls with array or object parameters (e.g. SetTodoList, ExitPlanMode, StrReplaceFile) fail with Pydantic validation errors:

Error validating JSON arguments: 1 validation error for Params
todos
  Input should be a valid list [type=list_type, input_value=[{"title": ...}], input_type=str]

The todos parameter is received as a JSON string instead of a parsed list/array.

Root Cause

In KimiToolset.handle() (kimi_cli/soul/toolset.py:301):

arguments: JsonType = json.loads(tool_call.function.arguments or "{}", strict=False)

The Moonshot API returns function.arguments with double-encoded JSON strings for nested array/object values. After json.loads parses the outer JSON, inner values like todos remain as strings:

API returns: {"todos": "[{\"title\": \"test\", \"status\": \"in_progress\"}]"}
                              ^--- string, not array!
json.loads → {"todos": "[{\"title\": \"test\", \"status\": \"in_progress\"}]"}
                        ^--- still a string!
Pydantic → ERROR: expected list, got str

Affected Tools

Reproduction

1. Call SetTodoList with todos=[{"title": "test", "status": "in_progress"}]
2. Observe Pydantic validation error

Proposed Fix

In kimi_cli/soul/toolset.py, KimiToolset.handle() method, replace:

arguments: JsonType = json.loads(tool_call.function.arguments or "{}", strict=False)

with:

raw = tool_call.function.arguments or "{}"
if isinstance(raw, dict):
    arguments: JsonType = raw
else:
    arguments = json.loads(raw, strict=False)
    # Moonshot API quirk: double-encoded JSON strings
    if isinstance(arguments, dict):
        for k, v in list(arguments.items()):
            if isinstance(v, str) and v and v[0] in ("[", "{"):
                try:
                    arguments[k] = json.loads(v, strict=False)
                except (json.JSONDecodeError, ValueError):
                    pass

Environment

• Kimi CLI: 1.46.0
• Pydantic: 2.12.5
• OpenAI SDK: 2.14.0
• Python: 3.14
• OS: Linux

Additional Context

The same issue exists in SimpleToolset.handle() in kosong/tooling/simple.py (the upstream library), but the Kimi CLI uses KimiToolset instead, which has its own handle() method with the same bug.

Start in kimi_cli/soul/toolset.py at KimiToolset.handle(), then reproduce the SetTodoList call described in the issue and inspect the parsed arguments before Pydantic validation. Compare the behavior with the stated Moonshot double-encoding case; done means array and object parameters reach the affected tools as parsed values without validation errors.