fix(kosong): recursively decode double-encoded tool-call arguments#2513
fix(kosong): recursively decode double-encoded tool-call arguments#2513nitishagar wants to merge 1 commit into
Conversation
Some providers (notably Moonshot) return function.arguments with nested array/object values as JSON strings. A single json.loads left them as strings, failing Pydantic validation on tools like SetTodoList/ExitPlanMode/ StrReplaceFile. Add a shared decode_tool_arguments helper in kosong.utils that recursively unwraps strings parsing to a dict or list (scalars and non-JSON strings are left untouched), and consume it from both SimpleToolset.handle and kimi-cli's KimiToolset.handle. Bumps kosong 0.55.0 -> 0.56.0 and the kimi-cli dep pin. Closes MoonshotAI#2406
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
| if isinstance(value, str): | ||
| try: | ||
| parsed = json.loads(value, strict=False) | ||
| except (json.JSONDecodeError, ValueError): | ||
| return value | ||
| if isinstance(parsed, (dict, list)): | ||
| return _unwrap(parsed) | ||
| return value |
There was a problem hiding this comment.
🔴 Writing or editing files whose contents are valid JSON now fails
Any string argument value whose text happens to be a valid JSON object or array is silently turned into structured data (_unwrap at packages/kosong/src/kosong/utils/json_args.py:37-44) before the tool runs, so a request to write or edit a file with JSON contents has its text replaced by a dict/list and is then rejected.
Impact: Common operations like writing package.json/tsconfig.json or replacing a JSON snippet in a file break with a validation error, even though the model sent perfectly valid input.
How the recursive unwrap corrupts string-typed parameters
The fix targets double-encoded params (e.g. todos returned as a JSON string). But _unwrap unconditionally re-parses ANY string that decodes to a dict or list, with no knowledge of the tool's declared parameter types. Consider WriteFile (src/kimi_cli/tools/file/write.py:29, content: str) writing a JSON file. The provider sends arguments = '{"path":"a.json","content":"{\"a\":1}"}'. The outer json.loads at packages/kosong/src/kosong/utils/json_args.py:60 correctly yields content as the string {"a":1}. Then _unwrap sees that string parses to a dict and replaces it with the dict {"a":1} (json_args.py:42-43).
When CallableTool2.call runs self.params.model_validate(arguments) (packages/kosong/src/kosong/tooling/__init__.py:296), pydantic's lax mode does NOT coerce a dict/list into a str field, so it raises ValidationError → ToolValidateError. The same applies to StrReplaceFile.Edit.old/new (src/kimi_cli/tools/file/replace.py:23-24) when replacing JSON text. Previously (single json.loads) these worked correctly.
Prompt for agents
The recursive _unwrap in packages/kosong/src/kosong/utils/json_args.py promotes ANY string that parses to a dict/list into that structured value, with no awareness of the tool parameter schema. This corrupts legitimate string-typed parameters whose content is valid JSON text. Concrete regressions: WriteFile.content (src/kimi_cli/tools/file/write.py, content: str) writing a JSON file (e.g. package.json) — the content string {"a":1} gets turned into a dict and then rejected by pydantic model_validate (packages/kosong/src/kosong/tooling/__init__.py:296) since dict cannot coerce to str; and StrReplaceFile.Edit.old/new (src/kimi_cli/tools/file/replace.py) replacing JSON snippets.
The double-encoding unwrap needs to be schema-aware rather than blindly promoting every JSON-looking string. Possible approaches: (1) only unwrap a string when the target field's declared type actually expects a list/dict (requires plumbing the pydantic params model into the decode step), or (2) perform the unwrap lazily only when initial validation fails and retry, or (3) restrict unwrapping so that string values are only promoted when validation would otherwise fail. Any fix must preserve the ability to write/edit files whose contents are valid JSON documents.
Was this helpful? React with 👍 or 👎 to provide feedback.
1 participant
Summary
function.argumentswith nested array/object values as JSON strings (double-encoding). After a singlejson.loads, values liketodosremained strings and failed Pydantic validation (Input should be a valid list).decode_tool_argumentshelper inkosong.utils.json_argsthat recursively unwraps any string that itself decodes to adictorlist. Scalars ("42","true") and non-JSON strings ("{oops}") are left unchanged, so genuine string fields are not corrupted.kosong.tooling.simple.SimpleToolset.handleandkimi_cli.soul.toolset.KimiToolset.handle.0.55.0→0.56.0and the kimi-cli dep pin; both changelogs updated under## Unreleased.Test plan
packages/kosong/tests/utils/test_json_args.pycovers E1–E8 (top-level and nested double-encoding, non-JSON strings preserved, scalar strings preserved, None/empty handling, normal passthrough, list-typed outer, mid-bracket values), outer-malformed re-raises, and two termination cases (bounded structural generators — no exponential string growth).tests/core/test_toolset.py::test_handle_unwraps_double_encoded_argumentsreproduces the reportedSetTodoList-style failure (stringtodos→ list) and asserts the tool now executes.Prior art
A prior PR #2407 attempted this but was closed without review. This PR supersedes it with a recursive (not single-level) unwrap, a shared helper, and the full edge-case suite.
Closes #2406; supersedes #2407 (closed).