feat(roundtable): 第二/三步模型自动容错(出问题自动换下一个模型)
模型出问题时按 config.yaml models[] 顺序自动切到下一个模型继续,并告知用户 「『X』模型异常 → 已切换『Y』继续」。前台实时 + 后台挂起两条路径都覆盖。 - 中间件兜底/熔断 AIMessage 打 LLM_ERROR_MARKER 标记,is_llm_error_message 可识别 - roundtable_model_fallback: fallback_model_chain 候选链 + reason_text + 总开关 env ROUNDTABLE_MODEL_FALLBACK - 触发: 异常(start_run失败/idle卡死) / 兜底错误文案 / 交付为空(仅seat/leader) - 后台 InProcessRoundtableGateway._run_turn_with_fallback 包住每轮 run - 前台 _special_run/_leader_run 收尾检测失败→model_switch 帧+换模型重流 - 结构化抽取 flow_extract round1 astream 同款换模型循环 - Step2 单席位异常隔离: 单席位失败不拖垮整批/整链/整个作业 - 时间线插「系统提示」气泡 + toast + 诊断 model_switch 事件 - Tests: test_roundtable_model_fallback.py Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
1 parent
ae2706dad3
commit
d383b0e9dd
18 files changed
+1033
-174
No files matched your search
@@ -101,6 +101,11 @@ export interface FlowExtractRequest {
|
||||
onValidating?: () => void;
|
||||
/** 校验/修复流式增量(同样含思考 + 正文,前端实时显示校验过程)。 */
|
||||
onValidateDelta?: (text: string) => void;
|
||||
/**
|
||||
* 模型自动容错:后端某模型抽取失败,已切到下一个模型重抽。上层据此**清空已累计的抽取增量**
|
||||
* (`onExtractDelta` 之前吐出的半截内容)并提示用户。`failedModel`/`nextModel`/`reason` 描述切换。
|
||||
*/
|
||||
onModelSwitch?: (info: { failedModel: string; nextModel: string; reason: string }) => void;
|
||||
}
|
||||
|
||||
/**
|
||||
@@ -161,6 +166,12 @@ export async function streamFlowExtract(req: FlowExtractRequest): Promise<FlowGr
|
||||
} else if (event === "validate_delta") {
|
||||
const c = typeof frame.content === "string" ? frame.content : "";
|
||||
if (c) req.onValidateDelta?.(c);
|
||||
} else if (event === "model_switch") {
|
||||
req.onModelSwitch?.({
|
||||
failedModel: typeof frame.failed_model === "string" ? frame.failed_model : "",
|
||||
nextModel: typeof frame.next_model === "string" ? frame.next_model : "",
|
||||
reason: typeof frame.reason === "string" ? frame.reason : "",
|
||||
});
|
||||
} else if (event === "final") {
|
||||
graph = (frame.graph as FlowGraphData) ?? null;
|
||||
} else if (event === "error") {
|
||||
|
||||
@@ -297,6 +297,27 @@ export interface StreamMultiAgentRequest {
|
||||
name?: string,
|
||||
args?: Record<string, unknown>,
|
||||
) => void;
|
||||
/**
|
||||
* 模型自动容错:后端检测到当前模型出问题(兜底错误文案 / 交付为空等),已自动切到下一个
|
||||
* 模型并即将重跑本轮时触发。上层据此把该角色「正在流式的失败气泡」重置(清空已累计的增量
|
||||
* 文本)并给用户一条「『X』模型异常,已切换『Y』继续」的提示。此回调可能多次触发(连续多个
|
||||
* 模型都失败时)。触发后,本次 streamMultiAgent 仍会继续用新模型流式,最终照常 resolve。
|
||||
*/
|
||||
onModelSwitch?: (info: ModelSwitchInfo) => void;
|
||||
}
|
||||
|
||||
/** 模型自动容错通知(`streamMultiAgent` 的 `onModelSwitch` 回调入参)。 */
|
||||
export interface ModelSwitchInfo {
|
||||
/** 角色:leader / seat / seat_parallel / report / summary / dashboard / ingest。 */
|
||||
role: string;
|
||||
/** 出问题的智能体 agent_name。 */
|
||||
agentName: string;
|
||||
/** 失败的模型名。 */
|
||||
failedModel: string;
|
||||
/** 已切换到的下一个模型名。 */
|
||||
nextModel: string;
|
||||
/** 失败原因(已翻译成简短中文,如「服务繁忙」)。 */
|
||||
reason: string;
|
||||
}
|
||||
|
||||
export interface ClarificationPayload {
|
||||
@@ -330,6 +351,8 @@ export interface StreamMultiAgentResult {
|
||||
content: string;
|
||||
/** Echoed agent_name of the agent that produced this run. */
|
||||
agentName: string | null;
|
||||
/** 实际产出本轮结果的模型(发生模型自动容错时即最终切换到的模型);后端未下发时为 null。 */
|
||||
modelUsed?: string | null;
|
||||
/**
|
||||
* Populated when the leader called ask_clarification — the run ended without
|
||||
* dispatching, and the model is asking the user a question instead. When
|
||||
@@ -555,6 +578,7 @@ export async function streamMultiAgent(
|
||||
content: "",
|
||||
agentName: null,
|
||||
clarification: null,
|
||||
modelUsed: null,
|
||||
};
|
||||
|
||||
// Inferred SSE phase + a tiny helper so we only fire onPhaseChange when the
|
||||
@@ -680,6 +704,28 @@ export async function streamMultiAgent(
|
||||
});
|
||||
}
|
||||
|
||||
// 模型自动容错帧:后端检测到当前模型出问题,已切到下一个模型并即将重跑本轮。
|
||||
// 不是终态帧(无 content)→ 不结束流。清空已累计的(失败尝试的)tool_call 增量,
|
||||
// 把 phase 退回 connected,并通知上层重置该角色气泡 + 给用户提示。随后新模型的
|
||||
// 流式增量会照常通过 onTextDelta 继续累计到(已被上层清空的)气泡。
|
||||
if (
|
||||
parsed &&
|
||||
typeof parsed === "object" &&
|
||||
(parsed as Record<string, unknown>).status === "model_switch"
|
||||
) {
|
||||
const obj = parsed as Record<string, unknown>;
|
||||
toolCallAcc.clear();
|
||||
currentPhase = "connected";
|
||||
req.onModelSwitch?.({
|
||||
role: typeof obj.role === "string" ? obj.role : "",
|
||||
agentName: typeof obj.agent_name === "string" ? obj.agent_name : req.agentName,
|
||||
failedModel: typeof obj.failed_model === "string" ? obj.failed_model : "",
|
||||
nextModel: typeof obj.next_model === "string" ? obj.next_model : "",
|
||||
reason: typeof obj.reason === "string" ? obj.reason : "",
|
||||
});
|
||||
continue;
|
||||
}
|
||||
|
||||
// Phase inference from LangGraph messages-tuple frames. We only flip
|
||||
// to tool_calling/tool_result on chunks that clearly indicate them;
|
||||
// everything else with non-empty AI text means we're streaming.
|
||||
@@ -780,6 +826,10 @@ export async function streamMultiAgent(
|
||||
if (typeof parsed.agent_name === "string") {
|
||||
result.agentName = parsed.agent_name;
|
||||
}
|
||||
{
|
||||
const mu = (parsed as Record<string, unknown>).model_used;
|
||||
if (typeof mu === "string" && mu) result.modelUsed = mu;
|
||||
}
|
||||
// SSE error 帧(后端 stream_generator 在席位/Step3 run 失败时下发
|
||||
// `{status:"error", content, agent_name}`):上报,便于收集用户侧报错。
|
||||
if (result.message === "error") {
|
||||
|
||||
Reference in new issue
Block a user