fix: 速度胶囊每轮只显示一次 + 技能调用按 run 合并统计
速度胶囊:原本每条 AI 消息都渲染,agentic 多轮 LLM 调用会出现多个胶囊。改为只在每轮 turn-final 消息(下一条 human 之前的最后一条 AI 消息)显示。 技能统计:tool_call_metrics 改为按 (run_id, skill_name) upsert 合并——同一 run 内同一技能的多次调用/重试合并为一行;任一成功即记成功,全部失败则记一行 error 并累积所有报错信息;非技能调用不受影响。 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
1 parent
4c38055d38
commit
028c5a34e6
5 files changed
+167
-17
No files matched your search
+124
-15
@@ -2,6 +2,7 @@
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import logging
|
||||
from collections.abc import AsyncIterator
|
||||
from datetime import UTC, datetime
|
||||
@@ -15,11 +16,99 @@ from deerflow.persistence.tool_metrics.model import ToolCallMetricRow
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
# Separator joining every failed attempt's message inside one merged skill row.
|
||||
_ERROR_SEP = "\n--- \n"
|
||||
# Hard cap on the accumulated ``error_message`` of a merged row.
|
||||
_MERGED_ERROR_LIMIT = 2000
|
||||
|
||||
# A single skill invocation often spans several tool calls (read SKILL.md, run
|
||||
# a script, ...), and the model may retry a skill within the same run. Those
|
||||
# fire-and-forget writes can interleave on the event loop, so the read → merge →
|
||||
# write of a skill row is guarded by a lock. ``asyncio.Lock`` binds to the loop
|
||||
# it is used on; the lock is therefore (re)created whenever the running loop
|
||||
# changes (the Gateway keeps one persistent loop, while the ``asyncio.run``
|
||||
# fallback path spins up a throwaway loop per call — already serialized).
|
||||
_merge_lock: asyncio.Lock | None = None
|
||||
_merge_lock_loop: object = None
|
||||
|
||||
|
||||
def _skill_merge_lock() -> asyncio.Lock:
|
||||
global _merge_lock, _merge_lock_loop
|
||||
loop = asyncio.get_running_loop()
|
||||
if _merge_lock is None or _merge_lock_loop is not loop:
|
||||
_merge_lock = asyncio.Lock()
|
||||
_merge_lock_loop = loop
|
||||
return _merge_lock
|
||||
|
||||
|
||||
def _normalize_created_at(value: datetime) -> datetime:
|
||||
return value if value.tzinfo else value.replace(tzinfo=UTC)
|
||||
|
||||
|
||||
def _merge_error_messages(existing: str | None, new: str | None) -> str | None:
|
||||
"""Append a failed attempt's message to the accumulated error text.
|
||||
|
||||
All failures are kept ("都统计下来"), separated and capped at
|
||||
:data:`_MERGED_ERROR_LIMIT` characters.
|
||||
"""
|
||||
new = (new or "").strip()
|
||||
if not new:
|
||||
return existing
|
||||
existing = (existing or "").strip()
|
||||
if not existing:
|
||||
return new
|
||||
if new in existing:
|
||||
# Identical retry failure — no point storing the same line twice.
|
||||
return existing
|
||||
combined = existing + _ERROR_SEP + new
|
||||
if len(combined) > _MERGED_ERROR_LIMIT:
|
||||
combined = combined[: _MERGED_ERROR_LIMIT - 3] + "..."
|
||||
return combined
|
||||
|
||||
|
||||
def _merge_skill_row(existing: ToolCallMetricRow, metric: ToolMetricRecord) -> None:
|
||||
"""Fold another same-run, same-skill call into an existing metric row.
|
||||
|
||||
Outcome rules:
|
||||
- any success in the run -> the row becomes a single ``success`` and earlier
|
||||
failures are dropped ("先报错、最后成功 -> 记一次成功");
|
||||
- all calls failed -> a single ``error`` row whose ``error_message``
|
||||
accumulates every attempt's message ("一直报错 -> 记一次,报错信息都统计下来").
|
||||
"""
|
||||
existing.duration_ms = (existing.duration_ms or 0) + (metric.duration_ms or 0)
|
||||
if metric.status == "success":
|
||||
if existing.status != "success":
|
||||
existing.status = "success"
|
||||
existing.error_category = None
|
||||
existing.error_message = None
|
||||
return
|
||||
# metric.status == "error"
|
||||
if existing.status == "success":
|
||||
# The skill already succeeded once this run — keep that outcome.
|
||||
return
|
||||
existing.status = "error"
|
||||
if metric.error_category:
|
||||
existing.error_category = metric.error_category
|
||||
existing.error_message = _merge_error_messages(existing.error_message, metric.error_message)
|
||||
|
||||
|
||||
def _row_from_metric(metric: ToolMetricRecord) -> ToolCallMetricRow:
|
||||
return ToolCallMetricRow(
|
||||
id=metric.id,
|
||||
created_at=_normalize_created_at(metric.created_at),
|
||||
user_id=metric.user_id,
|
||||
thread_id=metric.thread_id,
|
||||
run_id=metric.run_id,
|
||||
agent_name=metric.agent_name,
|
||||
tool_name=metric.tool_name,
|
||||
skill_name=metric.skill_name,
|
||||
duration_ms=metric.duration_ms,
|
||||
status=metric.status,
|
||||
error_category=metric.error_category,
|
||||
error_message=metric.error_message,
|
||||
)
|
||||
|
||||
|
||||
class SqlToolMetricsStore(ToolMetricsStore):
|
||||
"""SQLite/Postgres/MySQL-backed implementation; writes are fire-and-forget."""
|
||||
|
||||
@@ -27,28 +116,48 @@ class SqlToolMetricsStore(ToolMetricsStore):
|
||||
self._sf = session_factory
|
||||
|
||||
async def record(self, metric: ToolMetricRecord) -> None:
|
||||
row = ToolCallMetricRow(
|
||||
id=metric.id,
|
||||
created_at=_normalize_created_at(metric.created_at),
|
||||
user_id=metric.user_id,
|
||||
thread_id=metric.thread_id,
|
||||
run_id=metric.run_id,
|
||||
agent_name=metric.agent_name,
|
||||
tool_name=metric.tool_name,
|
||||
skill_name=metric.skill_name,
|
||||
duration_ms=metric.duration_ms,
|
||||
status=metric.status,
|
||||
error_category=metric.error_category,
|
||||
error_message=metric.error_message,
|
||||
)
|
||||
# Skill calls are merged per run: the model may touch one skill across
|
||||
# several tool calls / retries, and the admin wants a single row per
|
||||
# (run, skill) instead of one per call. Non-skill tool calls, and skill
|
||||
# calls with no run_id, fall back to a plain insert.
|
||||
if metric.skill_name and metric.run_id:
|
||||
await self._record_skill_merged(metric)
|
||||
return
|
||||
try:
|
||||
async with self._sf() as session:
|
||||
session.add(row)
|
||||
session.add(_row_from_metric(metric))
|
||||
await session.commit()
|
||||
except Exception:
|
||||
# Metrics are best-effort; never surface a write failure to the runtime.
|
||||
logger.warning("Failed to persist tool_call_metrics row %s", metric.id, exc_info=True)
|
||||
|
||||
async def _record_skill_merged(self, metric: ToolMetricRecord) -> None:
|
||||
"""Upsert a per-(run, skill) row, folding repeated calls into one."""
|
||||
try:
|
||||
async with _skill_merge_lock():
|
||||
async with self._sf() as session:
|
||||
existing = await session.scalar(
|
||||
select(ToolCallMetricRow)
|
||||
.where(
|
||||
ToolCallMetricRow.run_id == metric.run_id,
|
||||
ToolCallMetricRow.skill_name == metric.skill_name,
|
||||
)
|
||||
.order_by(ToolCallMetricRow.created_at)
|
||||
.limit(1)
|
||||
)
|
||||
if existing is None:
|
||||
session.add(_row_from_metric(metric))
|
||||
else:
|
||||
_merge_skill_row(existing, metric)
|
||||
await session.commit()
|
||||
except Exception:
|
||||
logger.warning(
|
||||
"Failed to merge tool_call_metrics for skill %s (run %s)",
|
||||
metric.skill_name,
|
||||
metric.run_id,
|
||||
exc_info=True,
|
||||
)
|
||||
|
||||
def _apply_filters(self, stmt, query: ToolMetricsQuery):
|
||||
if query.user_id:
|
||||
stmt = stmt.where(ToolCallMetricRow.user_id == query.user_id)
|
||||
|
||||
Reference in new issue
Block a user