feat(agent): add controlled task execution

This commit is contained in:
2026-09-24 04:19:02 +08:00
parent d018bcae67
commit b31033e9ce
48 changed files with 14498 additions and 86 deletions
+356 -35
View File
@@ -1,14 +1,20 @@
import { type Router } from "express";
import { randomUUID } from "node:crypto";
import {
UI_MESSAGE_STREAM_HEADERS,
type UIMessage,
type UIMessageChunk,
} from "ai";
import { z } from "zod";
import { createExecutionPlanSnapshot, executionPlanNeedsTaskPublish, readExecutionPlans } from "./executionPlan.js";
import { executionPlanSchema } from "./executionPlan.js";
import { isSupportedModel } from "../chat/models.js";
import { type ChatSessionBridge } from "../chat/sessionBridge.js";
import { getLocalAgentContext } from "../context/localContext.js";
import {
getLocalAgentContext,
getTrustedProductCredential,
} from "../context/localContext.js";
import { type FrontendActionCoordinator } from "../frontendAction/coordinator.js";
import { type LearningOrchestrator } from "../learning/orchestrator.js";
import { logger } from "../logger.js";
@@ -21,7 +27,14 @@ import {
import {
type SessionMetadataStore,
type SessionRecord,
readPurpose,
} from "../sessions/metadataStore.js";
import {
bindOperationalTaskCredential,
clearOperationalTaskLease,
operationalTaskClient,
releaseOperationalTaskCredential,
} from "../product/operational-task-client.js";
import { type SessionTranscriptStore } from "../sessions/transcriptStore.js";
import { type SessionUiStateStore } from "../sessions/uiStateStore.js";
import type { UIEnvelopePayload } from "../uiEnvelope/types.js";
@@ -60,6 +73,10 @@ import {
upsertBackendQuestion,
upsertBackendTodoUpdate,
} from "./chatUiState.js";
import {
agentTriggerContextSchema,
buildAgentTriggerPrompt,
} from "./agentTrigger.js";
const payloadSchema = z.object({
id: z.string().max(128).optional(),
@@ -71,6 +88,7 @@ const payloadSchema = z.object({
}).optional(),
approval_mode: z.enum(["request", "always"]).optional().default("request"),
capabilities: z.array(z.string().max(64)).max(20).optional().default([]),
trigger_context: agentTriggerContextSchema.optional(),
});
const toSessionUiStateContext = (sessionRecord: SessionRecord) => ({
@@ -121,6 +139,15 @@ const readLatestUserText = (messages: UIMessage[]) => {
const toUiMessages = (messages: unknown[] | undefined) =>
Array.isArray(messages) ? messages.filter(isUiMessage) : [];
export const toCollaborationMessages = (messages: UIMessage[]) =>
messages.flatMap((message) => {
if (message.role !== "user" && message.role !== "assistant") return [];
const textParts = message.parts.filter(
(part) => isObjectRecord(part) && part.type === "text" && typeof part.text === "string",
);
return textParts.length > 0 ? [{ ...message, parts: textParts } as UIMessage] : [];
});
const toLegacyMessages = (messages: UIMessage[]) =>
messages.flatMap((message) => {
const content = readUiMessageText(message);
@@ -142,6 +169,66 @@ const createFallbackUserUiMessage = (content: string): UIMessage => ({
parts: [{ type: "text", text: content }],
});
export const buildFailedTaskExecutionPlan = (history: Record<string, unknown>[]) => {
const previous = history.at(-1);
const parsed = previous ? executionPlanSchema.safeParse(previous) : null;
if (!parsed?.success) {
return {
title: "工况任务执行失败",
reason: "Agent 模型或工具执行失败,已记录失败并停止本次执行。",
nodes: [{
id: "AGENT_RUN_FAILURE",
title: "Agent 执行失败",
status: "failed" as const,
owner: "agent" as const,
detail: "模型或工具执行失败,等待人工检查或重新发起。",
evidence: [],
action_refs: [],
}],
edges: [],
};
}
let changed = false;
const nodes = parsed.data.nodes.map((node) => {
if (node.status === "running") {
changed = true;
return {
...node,
status: "failed" as const,
detail: "模型或工具执行失败,当前节点未完成,等待人工检查或重新发起。",
};
}
if (node.status === "pending" || node.status === "waiting") {
return {
...node,
status: "cancelled" as const,
detail: `${node.detail ? `${node.detail};` : ""}本次 Agent 执行失败,后续工作未启动。`,
};
}
return node;
});
if (!changed && !nodes.some((node) => node.status === "failed")) {
const known = new Set(nodes.map((node) => node.id));
let index = 1;
let id = "AGENT_RUN_FAILURE";
while (known.has(id)) id = `AGENT_RUN_FAILURE_${index++}`;
nodes.push({
id,
title: "Agent 执行失败",
status: "failed",
owner: "agent",
detail: "模型或工具执行失败,等待人工检查或重新发起。",
evidence: [],
action_refs: [],
});
}
return {
...parsed.data,
reason: "Agent 模型或工具执行失败,已记录失败并停止本次执行。",
nodes,
};
};
const pipeUiChunk = (
res: {
write: (chunk: string) => void;
@@ -182,6 +269,7 @@ export const registerChatStreamRoute = (
}: RegisterChatStreamRouteOptions,
) => {
const { activeRuns, lastRunStatuses } = runState;
const startingSessions = new Set<string>();
chatRouter.post("/stream", async (req, res) => {
const parsed = payloadSchema.safeParse(req.body);
if (!parsed.success) {
@@ -192,6 +280,8 @@ export const registerChatStreamRoute = (
return;
}
let boundOperationalSessionId: string | undefined;
let startingSessionId: string | undefined;
try {
const authContext = getLocalAgentContext(req);
const projectId = authContext.projectId;
@@ -199,15 +289,25 @@ export const registerChatStreamRoute = (
const userId = authContext.userId;
const actorKey = authContext.actorKey;
const projectKey = authContext.projectKey;
const trigger = parsed.data.trigger_context;
const requestUiMessages = toUiMessages(parsed.data.messages);
const promptText =
const requestPromptText =
parsed.data.message?.trim() || readLatestUserText(requestUiMessages);
const promptText = trigger
? buildAgentTriggerPrompt(trigger)
: requestPromptText;
if (!promptText) {
res.status(400).json({
message: "message or messages with latest user text is required",
});
return;
}
if (trigger && !parsed.data.session_id?.trim()) {
res.status(400).json({
message: "event trigger requires a reserved session_id",
});
return;
}
const requestedSessionId = resolveRequestedStreamSessionId({
sessionId: parsed.data.session_id,
promptText,
@@ -223,6 +323,19 @@ export const registerChatStreamRoute = (
res.status(404).json({ message: "session not found" });
return;
}
if (trigger && existingSessionRecord?.triggerId !== trigger.trigger_id) {
res.status(409).json({ message: "trigger does not match session" });
return;
}
if (
existingSessionRecord &&
readPurpose(existingSessionRecord) === "task_execution" &&
existingSessionRecord.modelId &&
parsed.data.model !== existingSessionRecord.modelId
) {
res.status(409).json({ message: "task execution model does not match the reserved session" });
return;
}
const hadExistingRuntimeSession = Boolean(existingSessionRecord);
const { binding, requestContext, created } = await sessionBridge.resolve({
@@ -237,6 +350,8 @@ export const registerChatStreamRoute = (
setRuntimeSessionContext({
...(getRuntimeSessionContext(binding.sessionId)!),
capabilities: parsed.data.capabilities,
purpose: existingSessionRecord ? readPurpose(existingSessionRecord) : "ordinary_chat",
taskId: existingSessionRecord?.taskId,
});
const { record: ensuredSessionRecord, created: sessionCreated } =
await sessionMetadataStore.ensure({
@@ -247,6 +362,31 @@ export const registerChatStreamRoute = (
userId,
});
const activeSessionRecord = await sessionMetadataStore.touch(ensuredSessionRecord);
const sessionPurpose = readPurpose(activeSessionRecord);
if (
activeRuns.get(activeSessionRecord.sessionId)?.status === "running" ||
startingSessions.has(activeSessionRecord.sessionId)
) {
res.status(409).json({
message: "session is already streaming",
session_id: activeSessionRecord.sessionId,
});
return;
}
startingSessions.add(activeSessionRecord.sessionId);
startingSessionId = activeSessionRecord.sessionId;
bindOperationalTaskCredential(
binding.sessionId,
getTrustedProductCredential(req),
);
boundOperationalSessionId = binding.sessionId;
const resolvedTaskContext = activeSessionRecord.taskId
&& (sessionPurpose === "task_collaboration" || sessionPurpose === "task_execution")
? await operationalTaskClient.getAgentContext(binding.sessionId, activeSessionRecord.taskId)
: null;
const taskContext = sessionPurpose === "task_collaboration"
? resolvedTaskContext
: null;
const hasRuntimeConversation = hadExistingRuntimeSession
? await runtimeHasConversation(runtime, binding.sessionId)
: false;
@@ -264,22 +404,22 @@ export const registerChatStreamRoute = (
toSessionUiStateContext(activeSessionRecord),
);
const persistedUiMessages = toUiMessages(initialSessionState?.messages);
let executionPlanHistory = initialSessionState?.executionPlans ?? readExecutionPlans(initialSessionState?.messages ?? []);
const visiblePersistedMessages = sessionPurpose === "task_collaboration"
? toCollaborationMessages(persistedUiMessages)
: persistedUiMessages;
const visibleRequestMessages = sessionPurpose === "task_collaboration"
? toCollaborationMessages(requestUiMessages)
: requestUiMessages;
const requestMessages =
requestUiMessages.length > 0
? requestUiMessages
visibleRequestMessages.length > 0
? visibleRequestMessages
: [
...persistedUiMessages,
...visiblePersistedMessages,
createFallbackUserUiMessage(promptText),
];
let latestUiMessages = requestMessages;
const baseMessages = toLegacyMessages(requestMessages);
if (activeRuns.get(activeSessionRecord.sessionId)?.status === "running") {
res.status(409).json({
message: "session is already streaming",
session_id: activeSessionRecord.sessionId,
});
return;
}
const recentTurns = await sessionTranscriptStore.getRecentTurns(historyContext, 8);
logger.info(
@@ -313,13 +453,17 @@ export const registerChatStreamRoute = (
const activeRun: ActiveRun = {
clientSessionId,
controller: abortController,
messages: initialMessages,
messages: updateLastAssistantMessage(initialMessages, (message) => ({
...message, executionPlans: executionPlanHistory,
})),
pendingPermissions: new Map(),
pendingQuestions: new Map(),
status: "running",
subscribers: new Set(),
};
activeRuns.set(clientSessionId, activeRun);
startingSessions.delete(activeSessionRecord.sessionId);
startingSessionId = undefined;
lastRunStatuses.set(clientSessionId, "running");
const sessionUiStateContext = toSessionUiStateContext(activeSessionRecord);
const assistantUiMessageId = `assistant-${Date.now().toString(36)}`;
@@ -426,15 +570,17 @@ export const registerChatStreamRoute = (
};
writeUiChunk({ type: "start", messageId: assistantUiMessageId });
let persistQueue = sessionUiStateStore.write(sessionUiStateContext, {
executionPlans: executionPlanHistory,
sessionId: activeSessionRecord.sessionId,
isTitleManuallyEdited: initialSessionState?.isTitleManuallyEdited ?? false,
messages: latestUiMessages,
messages: sessionPurpose === "task_collaboration" ? toCollaborationMessages(latestUiMessages) : latestUiMessages,
});
const queueSessionUiStatePersist = () => {
const snapshot = {
executionPlans: executionPlanHistory,
sessionId: activeSessionRecord.sessionId,
isTitleManuallyEdited: initialSessionState?.isTitleManuallyEdited ?? false,
messages: latestUiMessages,
messages: sessionPurpose === "task_collaboration" ? toCollaborationMessages(latestUiMessages) : latestUiMessages,
};
persistQueue = persistQueue
.catch((error) => {
@@ -486,9 +632,10 @@ export const registerChatStreamRoute = (
req.on("close", handleClientClose);
res.on("close", handleClientClose);
let taskArtifactPublishQueue: Promise<unknown> = Promise.resolve();
const publish = (event: string, data: Record<string, unknown>) => {
const subscriberData =
let subscriberData =
event === "token"
? {
...data,
@@ -610,6 +757,36 @@ export const registerChatStreamRoute = (
}),
);
}
} else if (event === "execution_plan") {
try {
const snapshot = createExecutionPlanSnapshot(data, [{ executionPlans: executionPlanHistory }], clientSessionId);
executionPlanHistory = [...executionPlanHistory, snapshot];
subscriberData = snapshot;
activeRun.messages = updateLastAssistantMessage(activeRun.messages, (message) => ({
...message,
executionPlans: [...(Array.isArray(message.executionPlans) ? message.executionPlans : []), snapshot],
}));
if (activeSessionRecord.taskId && executionPlanNeedsTaskPublish(data)) {
taskArtifactPublishQueue = taskArtifactPublishQueue
.catch(() => undefined)
.then(() => operationalTaskClient.publishArtifacts(binding.sessionId, activeSessionRecord.taskId!, {
requestId: randomUUID(),
sourceSessionId: clientSessionId,
executionPlan: snapshot,
}))
.catch((error) => {
logger.warn({ err: error, sessionId: clientSessionId, taskId: activeSessionRecord.taskId }, "failed to publish execution plan to task record");
});
}
} catch (error) {
logger.warn({ err: error, sessionId: clientSessionId }, "rejected invalid execution plan");
publish("progress", {
id: "execution-plan-rejected", phase: "execution_plan", status: "error",
title: "计划更新未通过校验,保留上一版",
detail: error instanceof Error ? error.message : "计划结构无效",
});
return;
}
} else if (event === "todo_update") {
const payload = data as TodoUpdatePayload;
activeRun.messages = updateLastAssistantMessage(activeRun.messages, (message) => ({
@@ -629,8 +806,17 @@ export const registerChatStreamRoute = (
}));
}
for (const subscriber of activeRun.subscribers) {
subscriber.write(event, subscriberData);
const hiddenCollaborationEvent = sessionPurpose === "task_collaboration" && [
"progress",
"todo_update",
"tool_call",
"ui_envelope",
"execution_plan",
].includes(event);
if (!hiddenCollaborationEvent) {
for (const subscriber of activeRun.subscribers) {
subscriber.write(event, subscriberData);
}
}
void queueSessionUiStatePersist().catch((error) => {
logger.warn({ err: error, sessionId: clientSessionId }, "failed to persist chat stream state");
@@ -640,19 +826,62 @@ export const registerChatStreamRoute = (
const unregisterFrontendActionSink = parsed.data.capabilities.includes("frontend-action@1")
? frontendActionCoordinator.registerSink(clientSessionId, (request) => publish("frontend_action", request as unknown as Record<string, unknown>))
: undefined;
let taskHeartbeat: ReturnType<typeof setInterval> | undefined;
let taskHeartbeatActive = true;
let taskHeartbeatPromise: Promise<void> | undefined;
let taskLeaseFailure: Error | undefined;
let taskRunFinalized = false;
const stopTaskHeartbeat = async () => {
taskHeartbeatActive = false;
if (taskHeartbeat) clearInterval(taskHeartbeat);
await taskHeartbeatPromise?.catch(() => undefined);
};
if (sessionPurpose === "task_execution" && activeSessionRecord.taskId) {
taskHeartbeat = setInterval(() => {
if (!taskHeartbeatActive || taskHeartbeatPromise) return;
const heartbeat = operationalTaskClient.heartbeatTask(
binding.sessionId,
activeSessionRecord.taskId!,
activeSessionRecord.runId!,
clientSessionId,
).then(() => undefined).catch((error) => {
if (!taskHeartbeatActive || abortController.signal.aborted) return;
taskLeaseFailure = error instanceof Error
? error
: new Error("task lease heartbeat failed");
taskHeartbeatActive = false;
if (taskHeartbeat) clearInterval(taskHeartbeat);
logger.warn(
{ err: error, sessionId: clientSessionId, taskId: activeSessionRecord.taskId },
"task lease heartbeat failed; stopping stale execution",
);
abortController.abort();
}).finally(() => {
if (taskHeartbeatPromise === heartbeat) taskHeartbeatPromise = undefined;
});
taskHeartbeatPromise = heartbeat;
}, 20_000);
}
taskHeartbeat?.unref();
try {
const preparedMessage = await buildPromptWithLearningContext(
memoryStore,
requestContext.actorKey,
requestContext.projectKey,
{
recentTurns,
persistedMessages: baseMessages,
message: promptText,
restoreConversation: shouldRestoreConversation,
},
);
const preparedMessage = taskContext
? [
"你正在与用户协作理解一个工况任务。只根据下面由后端授权返回的任务、计划、报告和证据快照解释判断;不要展示后台执行会话、工具轨迹或内部推理。若资料不足,明确说明缺口。用户要求变更任务时使用结构化工具写回。",
`任务快照:${JSON.stringify(taskContext)}`,
`用户问题:${promptText}`,
].join("\n\n")
: await buildPromptWithLearningContext(
memoryStore,
requestContext.actorKey,
requestContext.projectKey,
{
recentTurns,
persistedMessages: baseMessages,
message: promptText,
restoreConversation: shouldRestoreConversation,
},
);
const streamResult = await streamPromptResponse({
runtime,
sessionId: binding.sessionId,
@@ -668,6 +897,63 @@ export const registerChatStreamRoute = (
publish(event, data);
},
});
if (taskLeaseFailure) {
publish("error", {
session_id: clientSessionId,
message: "工况任务租约已失效,本次执行已停止,等待任务重新领取。",
detail: taskLeaseFailure.message,
});
} else if (streamResult.failed && sessionPurpose === "task_execution" && activeSessionRecord.taskId) {
publish("execution_plan", buildFailedTaskExecutionPlan(executionPlanHistory));
}
await taskArtifactPublishQueue;
await stopTaskHeartbeat();
if (
sessionPurpose === "task_execution"
&& activeSessionRecord.taskId
&& activeSessionRecord.runId
&& !taskLeaseFailure
) {
if (streamResult.failed || streamResult.aborted) {
await operationalTaskClient.failTaskRun(
binding.sessionId,
activeSessionRecord.taskId,
activeSessionRecord.runId,
clientSessionId,
streamResult.aborted ? "AGENT_ABORTED" : "AGENT_STREAM_FAILED",
streamResult.aborted ? "Agent execution was aborted" : "Agent stream failed",
);
taskRunFinalized = true;
} else {
try {
const completedTask = await operationalTaskClient.completeTaskRun(
binding.sessionId,
activeSessionRecord.taskId,
activeSessionRecord.runId,
clientSessionId,
);
taskRunFinalized = true;
if (completedTask.lifecycle_status === "failed") {
throw new Error("Server rejected the run result because required artifacts were incomplete");
}
} catch (error) {
if (!taskRunFinalized) {
const code = error && typeof error === "object" && "code" in error
&& typeof error.code === "string" ? error.code : "AGENT_COMPLETION_FAILED";
await operationalTaskClient.failTaskRun(
binding.sessionId,
activeSessionRecord.taskId,
activeSessionRecord.runId,
clientSessionId,
code,
error instanceof Error ? error.message : "Agent run completion failed",
);
taskRunFinalized = true;
}
throw error;
}
}
}
await persistQueue.catch((error) => {
logger.warn({ err: error, sessionId: clientSessionId }, "failed to persist chat stream state");
});
@@ -720,7 +1006,7 @@ export const registerChatStreamRoute = (
});
}
const latestTurn = extractLatestFrontendTurn(activeRun.messages);
if (latestTurn) {
if (latestTurn && sessionPurpose === "ordinary_chat") {
void learningOrchestrator.onTurnCompleted({
...latestTurn,
requestContext,
@@ -733,9 +1019,37 @@ export const registerChatStreamRoute = (
});
}
}
} catch (error) {
await stopTaskHeartbeat();
if (
sessionPurpose === "task_execution" &&
activeSessionRecord.taskId
) {
publish("execution_plan", buildFailedTaskExecutionPlan(executionPlanHistory));
await taskArtifactPublishQueue;
if (activeSessionRecord.runId && !taskLeaseFailure && !taskRunFinalized) {
await operationalTaskClient.failTaskRun(
binding.sessionId,
activeSessionRecord.taskId,
activeSessionRecord.runId,
clientSessionId,
abortController.signal.aborted ? "AGENT_ABORTED" : "AGENT_EXECUTION_ERROR",
abortController.signal.aborted
? "Agent execution was aborted"
: error instanceof Error ? error.message : "Agent execution failed",
).then(() => {
taskRunFinalized = true;
}).catch((finishError) => {
logger.warn({ err: finishError, sessionId: clientSessionId }, "failed to record task run failure");
});
}
}
throw error;
} finally {
await stopTaskHeartbeat();
await taskArtifactPublishQueue.catch(() => undefined);
unregisterFrontendActionSink?.();
if (abortController.signal.aborted) {
if (abortController.signal.aborted && !taskLeaseFailure) {
activeRun.messages = updateLastAssistantMessage(activeRun.messages, (message) => ({
...message,
content:
@@ -764,10 +1078,11 @@ export const registerChatStreamRoute = (
logger.warn({ err: error, sessionId: clientSessionId }, "failed to persist chat stream state");
});
sessionBridge.finalizeRequest(clientSessionId);
activeRun.status = abortController.signal.aborted
? activeRun.status === "aborted"
? "aborted"
: "aborted"
releaseOperationalTaskCredential(binding.sessionId);
activeRun.status = taskLeaseFailure
? "error"
: abortController.signal.aborted
? "aborted"
: activeRun.status === "running"
? "completed"
: activeRun.status;
@@ -791,6 +1106,12 @@ export const registerChatStreamRoute = (
res.end();
}
} catch (error) {
if (startingSessionId) {
startingSessions.delete(startingSessionId);
}
if (boundOperationalSessionId) {
releaseOperationalTaskCredential(boundOperationalSessionId);
}
const detail = error instanceof Error ? error.message : String(error);
logger.error({ err: error }, "chat stream failed");
if (res.headersSent) {