Files
deer-flow/backend/packages/harness/deerflow/agents/middlewares/token_usage_middleware.py
T
YuJitang eab7ae3d62 feat: stream subagent token usage to header via terminal task events (#2882)
* feat: real-time subagent token usage display in header and per-turn

Backend:
- Persist subagent token usage to AIMessage.usage_metadata via
  TokenUsageMiddleware, so accumulateUsage() naturally includes
  subagent tokens without frontend state management
- Cache subagent usage by tool_call_id in task_tool, write back
  to the dispatching AIMessage on next model response
- Emit subagent token usage on all terminal task events
  (task_completed, task_failed, task_cancelled, task_timed_out)
- Report subagent usage to parent RunJournal for API totals
- Search backward from ToolMessage to find dispatching AIMessage
  for correct multi-tool-call attribution

Frontend:
- Remove subagentUsage state, custom event handling, and prop
  threading — subagent tokens are now embedded in message metadata
- Simplify selectHeaderTokenUsage (no subagentUsage parameter)
- Per-turn inline badges show turn-specific usage via message
  accumulation
- Remove isLoading guard from MessageTokenUsageList for dynamic
  updates during streaming

* fix: prevent header token double counting from baseline reset race

onFinish, onError, and thread-switch useEffect all reset
pendingUsageBaselineMessageIdsRef to an empty Set. If
thread.isLoading is still true on the next render, all messages
pass the getMessagesAfterBaseline filter and their tokens are
added to backendUsage (which already includes them), causing
the header to display up to 2× the actual token count.

Capture current message IDs instead of using an empty Set so
that getMessagesAfterBaseline correctly returns no pending
messages even if thread.isLoading lags behind the stream end.

* fix: write back subagent tokens for all concurrent task tool calls

TokenUsageMiddleware only processed messages[-2], so when a
single model response dispatched multiple task tool calls only
the last ToolMessage had its cached subagent usage written back
to the dispatch AIMessage.usage_metadata. Earlier tasks' usage
stayed in _subagent_usage_cache indefinitely (leak) and never
appeared in the per-turn inline token display.

Walk backward through all consecutive ToolMessages before the
new AIMessage, and accumulate updates targeting the same
dispatch message into one state update so overlapping writes
don't clobber each other.

* fix: clean up subagent usage cache entry on task cancellation

When a task_tool invocation is cancelled via CancelledError, any
cached subagent usage entry leaked because the TokenUsageMiddleware
writeback path never fires after cancellation. Pop the cache entry
before re-raising to prevent unbounded growth of the module-level
_subagent_usage_cache dict.

* fix: address token usage review feedback

* fix: handle missing config for subagent usage cache

---------

Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
2026-05-13 23:52:19 +08:00

359 lines
13 KiB
Python

"""Middleware for logging token usage and annotating step attribution."""
from __future__ import annotations
import logging
from collections import defaultdict
from typing import Any, override
from langchain.agents import AgentState
from langchain.agents.middleware import AgentMiddleware
from langchain.agents.middleware.todo import Todo
from langchain_core.messages import AIMessage, ToolMessage
from langgraph.runtime import Runtime
logger = logging.getLogger(__name__)
TOKEN_USAGE_ATTRIBUTION_KEY = "token_usage_attribution"
def _string_arg(value: Any) -> str | None:
if isinstance(value, str):
normalized = value.strip()
return normalized or None
return None
def _normalize_todos(value: Any) -> list[Todo]:
if not isinstance(value, list):
return []
normalized: list[Todo] = []
for item in value:
if not isinstance(item, dict):
continue
todo: Todo = {}
content = _string_arg(item.get("content"))
status = item.get("status")
if content is not None:
todo["content"] = content
if status in {"pending", "in_progress", "completed"}:
todo["status"] = status
normalized.append(todo)
return normalized
def _todo_action_kind(previous: Todo | None, current: Todo) -> str:
status = current.get("status")
previous_content = previous.get("content") if previous else None
current_content = current.get("content")
if previous is None:
if status == "completed":
return "todo_complete"
if status == "in_progress":
return "todo_start"
return "todo_update"
if previous_content != current_content:
return "todo_update"
if status == "completed":
return "todo_complete"
if status == "in_progress":
return "todo_start"
return "todo_update"
def _build_todo_actions(previous_todos: list[Todo], next_todos: list[Todo]) -> list[dict[str, Any]]:
# This is the single source of truth for precise write_todos token
# attribution. The frontend intentionally falls back to a generic
# "Update to-do list" label when this metadata is missing or malformed.
previous_by_content: dict[str, list[tuple[int, Todo]]] = defaultdict(list)
matched_previous_indices: set[int] = set()
for index, todo in enumerate(previous_todos):
content = todo.get("content")
if isinstance(content, str) and content:
previous_by_content[content].append((index, todo))
actions: list[dict[str, Any]] = []
for index, todo in enumerate(next_todos):
content = todo.get("content")
if not isinstance(content, str) or not content:
continue
previous_match: Todo | None = None
content_matches = previous_by_content.get(content)
if content_matches:
while content_matches and content_matches[0][0] in matched_previous_indices:
content_matches.pop(0)
if content_matches:
previous_index, previous_match = content_matches.pop(0)
matched_previous_indices.add(previous_index)
if previous_match is None and index < len(previous_todos) and index not in matched_previous_indices:
previous_match = previous_todos[index]
matched_previous_indices.add(index)
if previous_match is not None:
previous_content = previous_match.get("content")
previous_status = previous_match.get("status")
if previous_content == content and previous_status == todo.get("status"):
continue
actions.append(
{
"kind": _todo_action_kind(previous_match, todo),
"content": content,
}
)
for index, todo in enumerate(previous_todos):
if index in matched_previous_indices:
continue
content = todo.get("content")
if not isinstance(content, str) or not content:
continue
actions.append(
{
"kind": "todo_remove",
"content": content,
}
)
return actions
def _describe_tool_call(tool_call: dict[str, Any], todos: list[Todo]) -> list[dict[str, Any]]:
name = _string_arg(tool_call.get("name")) or "unknown"
args = tool_call.get("args") if isinstance(tool_call.get("args"), dict) else {}
tool_call_id = _string_arg(tool_call.get("id"))
if name == "write_todos":
next_todos = _normalize_todos(args.get("todos"))
actions = _build_todo_actions(todos, next_todos)
if not actions:
return [
{
"kind": "tool",
"tool_name": name,
"tool_call_id": tool_call_id,
}
]
return [
{
**action,
"tool_call_id": tool_call_id,
}
for action in actions
]
if name == "task":
return [
{
"kind": "subagent",
"description": _string_arg(args.get("description")),
"subagent_type": _string_arg(args.get("subagent_type")),
"tool_call_id": tool_call_id,
}
]
if name in {"web_search", "image_search"}:
query = _string_arg(args.get("query"))
return [
{
"kind": "search",
"tool_name": name,
"query": query,
"tool_call_id": tool_call_id,
}
]
if name == "present_files":
return [
{
"kind": "present_files",
"tool_call_id": tool_call_id,
}
]
if name == "ask_clarification":
return [
{
"kind": "clarification",
"tool_call_id": tool_call_id,
}
]
return [
{
"kind": "tool",
"tool_name": name,
"description": _string_arg(args.get("description")),
"tool_call_id": tool_call_id,
}
]
def _infer_step_kind(message: AIMessage, actions: list[dict[str, Any]]) -> str:
if actions:
first_kind = actions[0].get("kind")
if len(actions) == 1 and first_kind in {"todo_start", "todo_complete", "todo_update", "todo_remove"}:
return "todo_update"
if len(actions) == 1 and first_kind == "subagent":
return "subagent_dispatch"
return "tool_batch"
if message.content:
return "final_answer"
return "thinking"
def _has_tool_call(message: AIMessage, tool_call_id: str) -> bool:
"""Return True if the AIMessage contains a tool_call with the given id."""
for tc in message.tool_calls or []:
if isinstance(tc, dict):
if tc.get("id") == tool_call_id:
return True
elif hasattr(tc, "id") and tc.id == tool_call_id:
return True
return False
def _build_attribution(message: AIMessage, todos: list[Todo]) -> dict[str, Any]:
tool_calls = getattr(message, "tool_calls", None) or []
actions: list[dict[str, Any]] = []
current_todos = list(todos)
for raw_tool_call in tool_calls:
if not isinstance(raw_tool_call, dict):
continue
described_actions = _describe_tool_call(raw_tool_call, current_todos)
actions.extend(described_actions)
if raw_tool_call.get("name") == "write_todos":
args = raw_tool_call.get("args") if isinstance(raw_tool_call.get("args"), dict) else {}
current_todos = _normalize_todos(args.get("todos"))
tool_call_ids: list[str] = []
for tool_call in tool_calls:
if not isinstance(tool_call, dict):
continue
tool_call_id = _string_arg(tool_call.get("id"))
if tool_call_id is not None:
tool_call_ids.append(tool_call_id)
return {
# Schema changes should remain additive where possible so older
# frontends can ignore unknown fields and fall back safely.
"version": 1,
"kind": _infer_step_kind(message, actions),
"shared_attribution": len(actions) > 1,
"tool_call_ids": tool_call_ids,
"actions": actions,
}
class TokenUsageMiddleware(AgentMiddleware):
"""Logs token usage from model responses and annotates the AI step."""
def _apply(self, state: AgentState) -> dict | None:
messages = state.get("messages", [])
if not messages:
return None
# Annotate subagent token usage onto the AIMessage that dispatched it.
# When a task tool completes, its usage is cached by tool_call_id. Detect
# the ToolMessage → search backward for the corresponding AIMessage → merge.
# Walk backward through consecutive ToolMessages before the new AIMessage
# so that multiple concurrent task tool calls all get their subagent tokens
# written back to the same dispatch message (merging into one update).
state_updates: dict[int, AIMessage] = {}
if len(messages) >= 2:
from deerflow.tools.builtins.task_tool import pop_cached_subagent_usage
idx = len(messages) - 2
while idx >= 0:
tool_msg = messages[idx]
if not isinstance(tool_msg, ToolMessage) or not tool_msg.tool_call_id:
break
subagent_usage = pop_cached_subagent_usage(tool_msg.tool_call_id)
if subagent_usage:
# Search backward from the ToolMessage to find the AIMessage
# that dispatched it. A single model response can dispatch
# multiple task tool calls, so we can't assume a fixed offset.
dispatch_idx = idx - 1
while dispatch_idx >= 0:
candidate = messages[dispatch_idx]
if isinstance(candidate, AIMessage) and _has_tool_call(candidate, tool_msg.tool_call_id):
# Accumulate into an existing update for the same
# AIMessage (multiple task calls in one response),
# or merge fresh from the original message.
existing_update = state_updates.get(dispatch_idx)
prev = existing_update.usage_metadata if existing_update else (getattr(candidate, "usage_metadata", None) or {})
merged = {
**prev,
"input_tokens": prev.get("input_tokens", 0) + subagent_usage["input_tokens"],
"output_tokens": prev.get("output_tokens", 0) + subagent_usage["output_tokens"],
"total_tokens": prev.get("total_tokens", 0) + subagent_usage["total_tokens"],
}
state_updates[dispatch_idx] = candidate.model_copy(update={"usage_metadata": merged})
break
dispatch_idx -= 1
idx -= 1
last = messages[-1]
if not isinstance(last, AIMessage):
if state_updates:
return {"messages": [state_updates[idx] for idx in sorted(state_updates)]}
return None
usage = getattr(last, "usage_metadata", None)
if usage:
input_token_details = usage.get("input_token_details") or {}
output_token_details = usage.get("output_token_details") or {}
detail_parts = []
if input_token_details:
detail_parts.append(f"input_token_details={input_token_details}")
if output_token_details:
detail_parts.append(f"output_token_details={output_token_details}")
detail_suffix = f" {' '.join(detail_parts)}" if detail_parts else ""
logger.info(
"LLM token usage: input=%s output=%s total=%s%s",
usage.get("input_tokens", "?"),
usage.get("output_tokens", "?"),
usage.get("total_tokens", "?"),
detail_suffix,
)
todos = state.get("todos") or []
attribution = _build_attribution(last, todos if isinstance(todos, list) else [])
additional_kwargs = dict(getattr(last, "additional_kwargs", {}) or {})
if additional_kwargs.get(TOKEN_USAGE_ATTRIBUTION_KEY) == attribution:
return {"messages": [state_updates[idx] for idx in sorted(state_updates)]} if state_updates else None
additional_kwargs[TOKEN_USAGE_ATTRIBUTION_KEY] = attribution
updated_msg = last.model_copy(update={"additional_kwargs": additional_kwargs})
state_updates[len(messages) - 1] = updated_msg
return {"messages": [state_updates[idx] for idx in sorted(state_updates)]}
@override
def after_model(self, state: AgentState, runtime: Runtime) -> dict | None:
return self._apply(state)
@override
async def aafter_model(self, state: AgentState, runtime: Runtime) -> dict | None:
return self._apply(state)