一次思考在轨迹里只落一条事件

思考流上线时漏了把它登记进 ephemeral 集合,导致每个流式增量都
落库——一条思考链在执行轨迹里被拆成几十条 llm.thinking 事件。

拆成两种事件:llm.thinking.delta 是流式增量,只推给画布做实时
滚动、不落库(与 llm.token 同类);llm.thinking 是调用收尾时落的
单条汇总,携带完整思考内容(截断 6000 字符,完整版在节点输出
工件里)。顺带修好刷新回放:此前回放重现不了思考(增量被隐藏、
汇总不存在),现在一条汇总事件直接恢复全文,时间线上也可展开查看。

验证:确定性假思考模型过完整管道(3 条增量实时推、汇总恰一条且
内容完整);真模型跑模板⑤,轨迹 29 条降到 18 条,llm.thinking
恰好 1 条。adaptive thinking 对简单任务自主跳过思考时两类事件
都不产生,已从工件核实是模型行为而非管道丢失。
This commit is contained in:
ThearchyHelios
2026-09-14 10:40:20 +08:00
parent e905426dfe
commit dcba9a63e0
5 changed files with 40 additions and 13 deletions
+3
View File
@@ -29,6 +29,9 @@ class EventType(StrEnum):
# 模型级
LLM_START = "llm.start"
LLM_TOKEN = "llm.token"
# 思考的流式增量:只推给画布做实时显示,不落库
LLM_THINKING_DELTA = "llm.thinking.delta"
# 一次完整思考:调用结束时落一条,这才是进轨迹的事件
LLM_THINKING = "llm.thinking"
LLM_END = "llm.end"
+25 -9
View File
@@ -116,6 +116,7 @@ async def _invoke_streaming(model: Any, messages: list[BaseMessage], ctx: NodeCo
"""
ctx.emit(EventType.LLM_START, message_count=len(messages))
chunks: list[Any] = []
response: AIMessage | None = None
try:
async for chunk in model.astream(messages):
chunks.append(chunk)
@@ -124,19 +125,34 @@ async def _invoke_streaming(model: Any, messages: list[BaseMessage], ctx: NodeCo
ctx.emit(EventType.LLM_TOKEN, delta=text)
reasoning = thinking_text(chunk)
if reasoning:
ctx.emit(EventType.LLM_THINKING, delta=reasoning)
# 增量只服务画布的实时滚动,不进轨迹(runner 标记为 ephemeral)
ctx.emit(EventType.LLM_THINKING_DELTA, delta=reasoning)
except NotImplementedError:
return await model.ainvoke(messages)
response = await model.ainvoke(messages)
except Exception as e: # noqa: BLE001
ctx.emit(EventType.LOG, level="warn", message=f"流式失败,回退非流式:{e}")
return await model.ainvoke(messages)
response = await model.ainvoke(messages)
if not chunks:
return await model.ainvoke(messages)
merged = chunks[0]
for chunk in chunks[1:]:
merged = merged + chunk
return merged
if response is None:
if not chunks:
response = await model.ainvoke(messages)
else:
merged = chunks[0]
for chunk in chunks[1:]:
merged = merged + chunk
response = merged
# 一次思考在轨迹里只占一条事件:调用收尾时落完整内容,
# 刷新页面回放时也靠它一次性恢复,而不是重放几十条增量
reasoning = thinking_text(response)
if reasoning:
ctx.emit(
EventType.LLM_THINKING,
text=reasoning[:6000],
chars=len(reasoning),
truncated=len(reasoning) > 6000,
)
return response
# --------------------------------------------------------------------------
+3 -2
View File
@@ -22,8 +22,9 @@ from app.engine.schema import GraphSpec, validate_graph
logger = logging.getLogger(__name__)
# token 事件量太大,只做实时推送不落库;回放时用节点的输出摘要重建时间线
_EPHEMERAL = {EventType.LLM_TOKEN}
# 流式增量(正文 token、思考 delta)量太大,只做实时推送不落库;
# 轨迹里正文靠 node.finished 的 preview,思考靠调用结束时的单条 llm.thinking 汇总
_EPHEMERAL = {EventType.LLM_TOKEN, EventType.LLM_THINKING_DELTA}
class RunManager:
+4 -1
View File
@@ -261,7 +261,7 @@ function ApprovalCard({ approval }: { approval: any }) {
// -------------------------------------------------------------------------
const HIDDEN_TYPES = new Set(['llm.token', 'llm.thinking'])
const HIDDEN_TYPES = new Set(['llm.token', 'llm.thinking.delta'])
function Timeline() {
const events = useStudio((s) => s.events)
@@ -359,6 +359,7 @@ const EVENT_META: Record<string, { text: string; tone: keyof typeof TONE }> = {
'node.skipped': { text: '跳过', tone: 'dim' },
'edge.taken': { text: '分支', tone: 'warn' },
'llm.start': { text: '调用模型', tone: 'dim' },
'llm.thinking': { text: '思考', tone: 'dim' },
'llm.end': { text: '模型返回', tone: 'dim' },
'tool.start': { text: '调用工具', tone: 'dim' },
'tool.end': { text: '工具完成', tone: 'ok' },
@@ -384,6 +385,8 @@ function describe(event: RunEvent): string {
case 'tool.end':
case 'tool.error':
return String(d.preview ?? '').slice(0, 300)
case 'llm.thinking':
return String(d.text ?? '').slice(0, 300)
case 'llm.end':
return `${d.model ?? ''} · ${d.input_tokens ?? 0}+${d.output_tokens ?? 0} tok${d.cost_usd ? ` · $${Number(d.cost_usd).toFixed(4)}` : ''}`
case 'sandbox.end':
+5 -1
View File
@@ -325,9 +325,13 @@ export const useStudio = create<StudioState>((set, get) => ({
case 'llm.token':
patch({ tokens: (runtime[nodeId!]?.tokens ?? '') + (event.data.delta ?? '') })
break
case 'llm.thinking':
case 'llm.thinking.delta':
patch({ thinking: (runtime[nodeId!]?.thinking ?? '') + (event.data.delta ?? '') })
break
case 'llm.thinking':
// 汇总事件:直接覆盖为完整思考。回放(刷新页面)时靠这一条恢复。
patch({ thinking: event.data.text ?? runtime[nodeId!]?.thinking })
break
case 'tool.start':
patch({
toolCalls: [...(runtime[nodeId!]?.toolCalls ?? []), { tool: event.data.tool, args: event.data.args }],