fix(messages): preserve non-string tool-call args in Hermes ShareGPT round-trip

[ebarkhordar]

Related Issue

Closes #4169

Description

HermesFunctionFormatter serialized tool calls by interpolating the dict into an f-string (Python repr) and parsed them back with json.loads(text.replace("'", '"')). That blanket quote swap corrupts any argument string containing an apostrophe (repr wraps such a value in double quotes, so the inner ' becomes a stray "), and True/False/None are not valid JSON. json.loads raised, the tool call was dropped, and from_sharegpt silently degraded the message to a plain BaseMessage with args lost. All-string args round-trip fine, which is why the existing test missed it.

Invariant: to_sharegpt() -> from_sharegpt() preserves func_name and args for any JSON-representable arguments.

Fix (single file):

• format_tool_call / format_tool_response emit json.dumps(..., ensure_ascii=False) instead of interpolating the dict.
• extract_tool_calls / extract_tool_response parse with json.loads, falling back to ast.literal_eval for legacy single-quoted (Python-repr) payloads so previously serialized data keeps round-tripping.

No dependency changes (ast and json are stdlib).

What is the purpose of this pull request?

• Bug fix

Verification

Verified in a clean python:3.12-slim container. The two new tests (test_hermes_sharegpt_roundtrip_preserves_non_string_args, test_hermes_extract_tool_calls_parses_legacy_repr_payload) fail on master and pass on this branch; the existing Hermes/ShareGPT tests stay green.

On master (source unchanged, new tests added):

Warning: Failed to parse tool call: Expecting ',' delimiter: line 1 column 49 (char 48)
FAILED test_hermes_sharegpt_roundtrip_preserves_non_string_args - assert isinstance(BaseMessage(...), FunctionCallingMessage)
FAILED test_hermes_extract_tool_calls_parses_legacy_repr_payload - assert 0 == 1
2 failed, 9 deselected

On this branch:

....                                                                     [100%]
4 passed, 7 deselected in 1.83s

The one unrelated failure in test/messages/test_func_message.py (test_roleplay_conversion_with_tools) is a missing OPENAI_API_KEY in the sandbox, not affected by this change.

Checklist

• I have read and agree to the AI-Generated Code Policy
• I have linked this PR to an issue (Hermes ShareGPT round-trip drops tool-call args with apostrophes, booleans or None #4169)
• Dependencies: none added or changed, no uv lock needed
• I have updated the tests accordingly
• Documentation: not needed
• Examples: not a new feature

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro

Run ID: 40a27854-3dcb-42af-9af6-66f70ffd4ea7

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[fengju0213]

thanks @ebarkhordar for opening the issue and pr！ lgtm

[ebarkhordar]

Thanks for the quick turnaround on both the issue and the PR. Glad the JSON round-trip fix was useful; I'll keep an eye on the ShareGPT conversion path.