[BUG] LangChain tool spans store raw RunnableConfig in metadata, crashing agentless trace encoding

[simonline]

Summary

LangChainIntegration._llmobs_set_meta_tags_from_tool stores the tool invocation config dict directly in span metadata without sanitization:

if tool_inputs.get("config"):
    metadata["tool_config"] = tool_inputs.get("config")

When the tool is executed by a LangGraph agent, this config is a RunnableConfig containing non-JSON-serializable objects, for example functools.partial (under configurable.__pregel_send) and AsyncCallbackManager (under callbacks). In agentless mode the JSON trace encoder then raises on flush and the entire trace payload is dropped:

TypeError: Object of type AsyncCallbackManager is not JSON serializable
  File ".../ddtrace/internal/encoding.py", line 211, in put
    encoded_trace = _json_dumps_bytes({"spans": [...]})

Depending on the graph wiring, the failing object can also be functools.partial. Every trace containing at least one LangGraph tool execution is lost, which makes LLM Observability silently blind for agent workloads in agentless setups.

Reproduction

• ddtrace 4.10.1, Python 3.12, langgraph + langchain-core current releases
• LLMObs.enable(agentless_enabled=True, ...) with the LangChain integration active
• Build any LangGraph StateGraph with a ToolNode and invoke it so a tool runs
• On span flush, the writer raises the TypeError above and drops the trace

Expected behavior

tool_config should go through the same safe_json sanitization that the integration already applies to tool arguments and I/O values in the neighbouring code paths.

Suggested fix

if tool_inputs.get("config"):
    metadata["tool_config"] = json.loads(safe_json(tool_inputs.get("config")))

Workaround

Wrapping _llmobs_set_meta_tags_from_tool to sanitize tool_inputs["config"] via safe_json before delegating to the original restores trace delivery.

[simonline]

Update: the suggested workaround is insufficient — the same crash occurs with other carriers, e.g. Pydantic model classes

We deployed the workaround from the issue description (wrapping _llmobs_set_meta_tags_from_tool to safe_json-sanitize tool_inputs["config"] before delegation) and still hit the identical failure, this time with:

TypeError: Object of type ModelMetaclass is not JSON serializable

Same traceback path (writer.py → _write_with_client → encoding.py:211 put → _json_dumps_bytes), triggered by a Pydantic model class reaching span data through an annotation path other than the tool config.

The underlying problem is broader than _llmobs_set_meta_tags_from_tool: JSONEncoderV2-based agentless encoding serializes each span's meta_struct — which carries the raw LLMObs span event including all annotated metadata — with a plain json.dumps and no default= fallback (ddtrace/internal/encoding.py, _item_to_dict / _json_dumps_bytes). Any single non-serializable object anywhere in any span (callback managers, functools.partial, Pydantic model classes, …) therefore drops the entire trace payload.

Suggestion: in addition to sanitizing tool_config/tool_info at the integration layer, harden the encoder itself, e.g.:

def _json_dumps_bytes(obj) -> bytes:
    return json.dumps(obj, default=repr).encode("utf-8", errors="backslashreplace")

so a single bad value degrades to its repr instead of silently killing the whole payload. We're running exactly this as a monkeypatch and it resolves both the AsyncCallbackManager and the ModelMetaclass variants.

Environment: ddtrace 4.10.1, Python 3.12, langchain-core 0.3.80, langgraph 0.3.31, pydantic 2.13.4, LLMObs agentless enabled.