Submission checklist
Package (Required)
Feature Description
The on_tool_start stream event currently exposes name, tags, run_id, metadata, parent_ids, and data.input — but not tool_call_id.
LangChain already tracks tool_call_id for each tool run: BaseTool.arun passes it to the callback manager, _AstreamEventsCallbackHandler stores it in run_info during on_tool_start (since #33731), and it's surfaced on on_tool_error.data (also #33731). The data is right there; it's just not emitted on on_tool_start.
import asyncio
from langchain.agents import create_agent
from langchain_core.tools import tool
@tool
def slow_lookup(query: str) -> str:
"""A tool that takes a bit of time so we can observe its start event."""
return f"result for {query}"
agent = create_agent(model=..., tools=[slow_lookup])
async def demonstrate_issue():
async for event in agent.astream_events(
{"messages": "look up alpha and beta in parallel"},
version="v2",
):
if event["event"] == "on_tool_start":
# event["data"] has "input" but no "tool_call_id"
assert "tool_call_id" not in event["data"]
asyncio.run(demonstrate_issue())
### Use Case
Streaming UI consumers building per-tool-call panels need to attribute every on_tool_* event to the originating tool call as soon as the tool starts — not wait for on_tool_end / on_tool_error. This matters especially under parallel tool calls of the same tool (e.g., a supervisor agent dispatching several sub-agents of the same kind in one assistant turn), where name alone isn't sufficient to disambiguate which streamed events belong to which call, and falling back to run_id/parent traversal is fragile.
on_tool_error already exposes tool_call_id in its event data (added in #33731). Doing the same for on_tool_start is the natural parity.
### Proposed Solution
Mirror the on_tool_error change from #33731 on the on_tool_start send:
```python
# libs/core/langchain_core/tracers/event_stream.py — in on_tool_start
self._send(
{
"event": "on_tool_start",
"data": {
"input": inputs or {},
"tool_call_id": kwargs.get("tool_call_id"), # ← new
},
"name": name_,
"tags": tags or [],
"run_id": str(run_id),
"metadata": metadata or {},
"parent_ids": self._get_parent_ids(run_id),
},
"tool",
)The kwarg is already passed in (it's used at line 311 to populate run_info);
this just surfaces it on the emitted event too.
Alternatives Considered
- Reading the streamed metadata field. Possible if the BE manually
splices tool_call_id into the tool's runtime config metadata via a
BaseTool subclass that overrides ainvoke. Works, but every downstream
agent author has to invent the same workaround. The data should be
first-class on the event.
- Walking parent_ids / run_id to recover the originating tool call.
Fragile — LangGraph collapses parent relationships, and run_id is
separate from tool_call_id.
Additional Context
Parallel to #33597 / #33731. Same justification (stateless agents, streaming-UI tool-call mapping), same one-line surface fix.
Submission checklist
Package (Required)
Feature Description
The
on_tool_startstream event currently exposesname,tags,run_id,metadata,parent_ids, anddata.input— but nottool_call_id.LangChain already tracks
tool_call_idfor each tool run:BaseTool.arunpasses it to the callback manager,_AstreamEventsCallbackHandlerstores it inrun_infoduringon_tool_start(since #33731), and it's surfaced onon_tool_error.data(also #33731). The data is right there; it's just not emitted onon_tool_start.The kwarg is already passed in (it's used at line 311 to populate run_info);
this just surfaces it on the emitted event too.
Alternatives Considered
splices tool_call_id into the tool's runtime config metadata via a
BaseTool subclass that overrides ainvoke. Works, but every downstream
agent author has to invent the same workaround. The data should be
first-class on the event.
Fragile — LangGraph collapses parent relationships, and run_id is
separate from tool_call_id.
Additional Context
Parallel to #33597 / #33731. Same justification (stateless agents, streaming-UI tool-call mapping), same one-line surface fix.