@gt-fe/eap-runtime 0.0.21 → 0.0.22

package/dist/runtime/agent-runtime.js

@@ -7,20 +7,20 @@
  * @see 04-runtime-core/design.md §1.5 / §1.7
  */
 import { randomUUID } from 'node:crypto';
-import { HITL_TIMEOUT_ERROR_CODE, HITL_TIMEOUT_MESSAGE, HitlAlreadyTimedOutError, HitlNotWaitingError, buildHitlWaitedPayload, isHitlDeadlinePassed, readPersistedHitlWaiting, unwrapHitlPaused, } from '../graph/hitl/hitl-handler.js';
+import { HITL_TIMEOUT_ERROR_CODE, HITL_TIMEOUT_MESSAGE, HitlAlreadyTimedOutError, HitlNotWaitingError, buildHitlWaitedPayload, isHitlDeadlinePassed, loadHitlCheckpoint, readPersistedHitlWaiting, unwrapHitlPaused, } from '../graph/hitl/hitl-handler.js';
 import { resolveAgentSystemPromptFromLoader } from '../registry/graph-loader.js';
 import { toLoadedSkillMap } from '../tool/system-prompt-builder.js';
 import { assembleLoadedTools } from '../tool/harness-primitives.js';
 import { createFileLlmTurnLogFromEnv, } from '../debug/llm-turn-log.js';
 import { buildHitlWaitingTaskOutput, buildTurnGraphMessages, buildTurnTranscriptFromStreamEvents, extractHitlThreadId, extractStreamReasoning, applyStreamTextDelta, createStreamAssistantBuckets, visibleStreamAssistantText, streamAssistantTexts, messageRecordsToGraphMessages, persistHitlWaitingPlaceholder, persistTurnTranscript, persistUserTurnMessage, resolveCheckpointThreadId, appendTerminalToolResult, } from '../session/persist-turn-messages.js';
-import { iterSessionCompression, } from '../session/apply-session-compression.js';
+import { iterSessionCompression, withCompressionSummaryModel, } from '../session/apply-session-compression.js';
 import { applySessionMemory } from '../session/apply-session-memory.js';
 import { resolveTokenDetail, shouldEmitTokenSse } from '../token/token-detail.js';
 import { AppError, ErrorCodes } from '@gt-fe/eap-contracts';
 import { resolveExecutionContext } from '../governance/governance-client.js';
 import { resolveAgentCode } from '../common/agent-code.js';
 import { resolveTraceId } from '../common/trace-id.js';
-import { collectStreamNodeTokenUsage, recordTaskTokenUsages, tokenUsagesFromLlmChannel, } from '../token/record-task-token-usage.js';
+import { collectStreamNodeTokenUsage, recordTaskTokenUsages, summarizeTokenUsages, tokenUsagesFromLlmChannel, } from '../token/record-task-token-usage.js';
 import { resolveSkillInputFromInvokeInput } from '../registry/skill-io-guard.js';
 const HITL_TERMINAL_STATUSES = [
     'success',
@@ -280,15 +280,18 @@ export class AgentRuntime {
             : undefined;
         const agentSystemPrompt = await resolveAgentSystemPromptFromLoader(this.deps.graphLoader, agentCode, params.agentVersion);
         let history = messageRecordsToGraphMessages(sessionContext.messages);
+        const compressionOptions = await withCompressionSummaryModel({
+            ...this.deps.sessionCompression,
+            turnLog: this.deps.sessionCompression?.turnLog ?? this.deps.turnLog,
+        }, () => this.deps.graphLoader
+            ? this.deps.graphLoader.resolveDefaultModel(agentCode, params.agentVersion)
+            : Promise.resolve(undefined));
         for await (const step of iterSessionCompression({
             history,
             sessionId,
             userId,
             taskId: task.id,
-            options: {
-                ...this.deps.sessionCompression,
-                turnLog: this.deps.sessionCompression?.turnLog ?? this.deps.turnLog,
-            },
+            options: compressionOptions,
         })) {
             if (step.kind === 'history') {
                 history = step.messages;
@@ -367,6 +370,7 @@ export class AgentRuntime {
             }
             await recordTaskTokenUsages(this.deps.taskService, task.id, nodeTokenUsages);
             await this.deps.taskService.updateStatus(task.id, 'success');
+            const tokenSummary = summarizeTokenUsages(nodeTokenUsages);
             await persistTurnTranscript(this.deps.messageService, {
                 sessionId,
                 taskId: task.id,
@@ -380,7 +384,12 @@ export class AgentRuntime {
                     toolCallReasoningContent,
                     assistantReasoningContent,
                 }),
+                assistantTokenCount: tokenSummary.total.output,
+                assistantDurationMs: Date.now() - startTime,
             });
+            if (shouldEmitTokenSse(tokenDetail, 'stats.token_usage')) {
+                yield this.toSSE('stats.token_usage', correlationId, task.id, tokenSummary);
+            }
             yield this.toSSE('task.completed', correlationId, task.id, {
                 taskId: task.id,
                 duration: Date.now() - startTime,
@@ -559,6 +568,7 @@ export class AgentRuntime {
             }
             await recordTaskTokenUsages(this.deps.taskService, taskId, nodeTokenUsages);
             await this.deps.taskService.updateStatus(taskId, 'success', visibleStreamAssistantText(assistantBuckets) || undefined);
+            const tokenSummary = summarizeTokenUsages(nodeTokenUsages);
             if (sessionId) {
                 await persistTurnTranscript(this.deps.messageService, {
                     sessionId,
@@ -573,8 +583,13 @@ export class AgentRuntime {
                         toolCallReasoningContent,
                         assistantReasoningContent,
                     }),
+                    assistantTokenCount: tokenSummary.total.output,
+                    assistantDurationMs: Date.now() - startTime,
                 });
             }
+            if (shouldEmitTokenSse(tokenDetail, 'stats.token_usage')) {
+                yield this.toSSE('stats.token_usage', correlationId, taskId, tokenSummary);
+            }
             yield this.toSSE('task.completed', correlationId, taskId, {
                 taskId,
                 duration: Date.now() - startTime,
@@ -778,16 +793,18 @@ export class AgentRuntime {
         const sessionId = typeof task.sessionId === 'string' && task.sessionId.trim()
             ? task.sessionId.trim()
             : undefined;
-        const threadId = waiting?.threadId ??
+        const hinted = waiting?.threadId ??
             (typeof response.threadId === 'string' && response.threadId.trim()
                 ? response.threadId.trim()
                 : undefined) ??
             extractHitlThreadId(task.output) ??
             sessionId;
-        if (!threadId) {
+        if (!hinted) {
             throw new AppError(ErrorCodes.TASK_HITL_INVALID_STATE, `Task ${taskId} 缺少 sessionId/hitlThreadId，无法冷恢复 HITL Checkpoint`, { exposeMessage: true });
         }
-        const checkpoint = await this.deps.checkpointSaver.load(threadId);
+        const resolved = await loadHitlCheckpoint(this.deps.checkpointSaver, hinted);
+        const threadId = resolved.threadId;
+        const checkpoint = resolved.checkpoint;
         const persisted = readPersistedHitlWaiting(checkpoint?.state);
         const behavior = persisted?.config.timeoutBehavior ?? 'error';
         if (persisted &&

package/dist/runtime/dev-runtime.d.ts

@@ -84,6 +84,7 @@ export interface DevRuntime {
  * 装配本地 dev 模式所需的全部 Runtime 依赖
  *
  * @returns DevRuntime（含 invokeOrchestrator / streamOrchestrator / 诊断摘要）
- * @throws EAP_LLM_MODE=local 但解不出任何密钥时抛错，避免静默 Echo
+ * @throws `EAP_LLM_MODE=local` 且解不出密钥
+ * @throws 已配 `FOUNDATION_SERVICE_URL` 且 `EAP_LLM_CONNECTIONS` 不是合法 JSON 对象
  */
 export declare function createDevRuntime(options?: CreateDevRuntimeOptions): DevRuntime;

package/dist/runtime/dev-runtime.js

@@ -25,7 +25,7 @@
  */
 import { createDefaultLLMProvider } from '../providers/echo-llm-provider.js';
 import { HttpLLMProvider } from '../providers/http-llm-provider.js';
-import { createLocalLlmProvider, requireLocalLlmConnections } from '../providers/local-llm-provider.js';
+import { createLocalLlmProvider, createOptionalLocalLlmProvider, requireLocalLlmConnections, } from '../providers/local-llm-provider.js';
 import { RoutingLLMProvider } from '../providers/routing-llm-provider.js';
 import { createToolExecutor } from '../tool/create-tool-executor.js';
 import { DevToolExecutor } from '../tool/dev-tool-executor.js';
@@ -44,7 +44,8 @@ import { isRuntimeMemoryEnabled } from '../session/apply-session-memory.js';
  * 装配本地 dev 模式所需的全部 Runtime 依赖
  *
  * @returns DevRuntime（含 invokeOrchestrator / streamOrchestrator / 诊断摘要）
- * @throws EAP_LLM_MODE=local 但解不出任何密钥时抛错，避免静默 Echo
+ * @throws `EAP_LLM_MODE=local` 且解不出密钥
+ * @throws 已配 `FOUNDATION_SERVICE_URL` 且 `EAP_LLM_CONNECTIONS` 不是合法 JSON 对象
  */
 export function createDevRuntime(options) {
     const foundationUrl = process.env.FOUNDATION_SERVICE_URL ?? 'http://localhost:3100';
@@ -72,7 +73,12 @@ export function createDevRuntime(options) {
         llmProviderKind = 'openai-http';
     }
     else if (foundationConfigured) {
-        llmProvider = new RoutingLLMProvider({ foundationUrl, turnLog });
+        const fallbackProvider = createOptionalLocalLlmProvider(turnLog);
+        llmProvider = new RoutingLLMProvider({
+            foundationUrl,
+            turnLog,
+            ...(fallbackProvider ? { fallbackProvider } : {}),
+        });
         llmProviderKind = 'routing';
     }
     else {

package/dist/session/apply-session-compression.d.ts

@@ -23,6 +23,8 @@ export interface SessionCompressionOptions {
     compressionRepository?: ICompressionRepository;
     maxRounds?: number;
     preserveRecentRounds?: number;
+    /** 摘要模型；与图节点同一套：Agent `models.default`。不要传字符串 `default`。 */
+    summaryModel?: string;
 }
 export interface RunSessionCompressionParams {
     history: GraphMessage[];
@@ -72,6 +74,8 @@ export declare function formatCompressionCompletedMessage(input: {
     error?: string;
 }): string;
 export declare function applySessionCompression(params: RunSessionCompressionParams): Promise<GraphMessage[]>;
+/** 未显式 summaryModel 时写入 Agent `models.default`。压缩未开则不解析。 */
+export declare function withCompressionSummaryModel(options: SessionCompressionOptions | undefined, resolveDefaultModel: () => Promise<string | undefined>, env?: NodeJS.ProcessEnv): Promise<SessionCompressionOptions | undefined>;
 /**
  * 流式入口用：先产出 started，再跑摘要，再产出 completed，最后给出本轮历史。
  * SDK `stream` 可在 started 之后才继续 next()，HTTP SSE 可立刻 emit。

package/dist/session/apply-session-compression.js

@@ -49,6 +49,19 @@ export async function applySessionCompression(params) {
     const result = await runSessionCompression(params);
     return result.messages;
 }
+/** 未显式 summaryModel 时写入 Agent `models.default`。压缩未开则不解析。 */
+export async function withCompressionSummaryModel(options, resolveDefaultModel, env = process.env) {
+    if (!options)
+        return options;
+    if (!isRuntimeCompressionEnabled(options.enabled, env))
+        return options;
+    if (options.summaryModel?.trim())
+        return options;
+    const summaryModel = (await resolveDefaultModel())?.trim();
+    if (!summaryModel || summaryModel.toLowerCase() === 'default')
+        return options;
+    return { ...options, summaryModel };
+}
 /**
  * 流式入口用：先产出 started，再跑摘要，再产出 completed，最后给出本轮历史。
  * SDK `stream` 可在 started 之后才继续 next()，HTTP SSE 可立刻 emit。
@@ -151,7 +164,7 @@ async function runDecidedSessionCompression(params, decision) {
             maxRounds,
             preserveRecentRounds: options.preserveRecentRounds ?? DEFAULT_COMPRESSION_CONFIG.preserveRecentRounds,
             ...(decision.window ? { maxContextTokens: decision.window } : {}),
-        }, async (msgs) => summarizeWithProvider(provider, msgs));
+        }, async (msgs) => summarizeWithProvider(provider, msgs, options.summaryModel));
         const compressed = await compression.compress(history);
         if (compressed === history) {
             return {
@@ -242,18 +255,24 @@ async function lastTurnUsage(turnLog, sessionId) {
         return undefined;
     return { used: last.used, window: last.window };
 }
-async function summarizeWithProvider(provider, messages) {
+async function summarizeWithProvider(provider, messages, summaryModel) {
     const body = messages
         .map((m) => `${m.role}: ${m.content}`)
         .join('\n')
         .slice(0, 12000);
+    const raw = summaryModel?.trim();
+    const model = raw && raw.toLowerCase() !== 'default' ? raw : '';
     const response = await provider.invoke([
         {
             role: 'system',
             content: '将对话历史压成短摘要，保留用户需求、结论和未完成事项。',
         },
         { role: 'user', content: body },
-    ], { model: 'default', temperature: 0.2, maxTokens: 1024 });
+    ], {
+        model,
+        temperature: 0.2,
+        maxTokens: 1024,
+    });
     const text = response.content?.trim();
     if (!text) {
         throw new Error('empty compression summary');

package/dist/session/index.d.ts

@@ -10,6 +10,6 @@ export type { FileMessageRepositoryOptions } from './file-message.repository.js'
 export { SessionContextService } from './session-context.service.js';
 export type { ResolveSessionParams, SessionContext, } from './session-context.service.js';
 export { CompressionMiddleware, createCompressionMiddleware } from './compression-middleware.js';
-export { applySessionCompression, runSessionCompression, decideSessionCompression, iterSessionCompression, formatCompressionStartedMessage, formatCompressionCompletedMessage, isRuntimeCompressionEnabled, isRuntimeCompressionPersist, type SessionCompressionOptions, type RunSessionCompressionParams, type RunSessionCompressionResult, type SessionCompressionDecision, type SessionCompressionStep, } from './apply-session-compression.js';
+export { applySessionCompression, withCompressionSummaryModel, runSessionCompression, decideSessionCompression, iterSessionCompression, formatCompressionStartedMessage, formatCompressionCompletedMessage, isRuntimeCompressionEnabled, isRuntimeCompressionPersist, type SessionCompressionOptions, type RunSessionCompressionParams, type RunSessionCompressionResult, type SessionCompressionDecision, type SessionCompressionStep, } from './apply-session-compression.js';
 export type { CompressionMiddlewareConfig } from './compression-middleware.js';
 export { applySessionMemory, isRuntimeMemoryEnabled, type SessionMemoryOptions, } from './apply-session-memory.js';

package/dist/session/index.js

@@ -10,5 +10,5 @@ export { FileMessageRepository } from './file-message.repository.js';
 export { SessionContextService } from './session-context.service.js';
 // 压缩中间件
 export { CompressionMiddleware, createCompressionMiddleware } from './compression-middleware.js';
-export { applySessionCompression, runSessionCompression, decideSessionCompression, iterSessionCompression, formatCompressionStartedMessage, formatCompressionCompletedMessage, isRuntimeCompressionEnabled, isRuntimeCompressionPersist, } from './apply-session-compression.js';
+export { applySessionCompression, withCompressionSummaryModel, runSessionCompression, decideSessionCompression, iterSessionCompression, formatCompressionStartedMessage, formatCompressionCompletedMessage, isRuntimeCompressionEnabled, isRuntimeCompressionPersist, } from './apply-session-compression.js';
 export { applySessionMemory, isRuntimeMemoryEnabled, } from './apply-session-memory.js';

package/dist/session/persist-turn-messages.d.ts

@@ -21,6 +21,7 @@ export declare function persistAssistantTurnMessage(messageService: MessageServi
     toolCalls?: unknown[];
     /** DeepSeek thinking：多轮回传用 */
     reasoningContent?: string;
+    durationMs?: number | null;
 }): Promise<void>;
 /**
  * HITL 等待态占位消息：避免早退时本轮只有 user、历史偏薄
@@ -80,6 +81,7 @@ export declare function sessionVisibleAssistantTexts(texts: Array<{
  * invoke 路径图增量。tool_calls / tool 行原样保留，纯文本助手收成一条放在末尾（与 stream 落库顺序一致）。
  * GraphMessage 无 nodeId：`finishPoints` 只含一个实际节点 id 时取最后一条纯文本，不是按 id 筛选（官方 DAG 的 `merge` 在末尾）。
  * 没有实际节点 id 或仅为 `__end__` 时把多段正文用空行拼进这一条。
+ * 无 toolCallId 的 tool 行不进 Session（code 节点结果）。
  */
 export declare function sessionVisibleGraphMessages(messages: GraphMessage[], finishPoints?: string[]): GraphMessage[];
 /**
@@ -138,8 +140,10 @@ export declare function buildHitlWaitingTaskOutput(params: {
 /** 从 Task.output 解析 Checkpoint thread_id（含嵌套 hitl.threadId） */
 export declare function extractHitlThreadId(output: unknown): string | undefined;
 /**
- * Checkpoint thread_id 契约：HITL/多轮优先且应唯一等于 sessionId；
+ * Checkpoint 父图 thread_id：HITL/多轮优先且应唯一等于 sessionId；
  * 无 session 时退到 `task-${taskId}`。勿用 correlationId 作为 Checkpoint key。
+ * 子图槽位再经 `checkpointStorageThreadId`（`thread_id#父节点id`），
+ * 与 LangGraph put / 子图内 HITL interrupt 同一把钥匙。
  *
  * 注意：session 级 thread + 全量 Session 历史同时注入会导致 messages 双重 append。
  * 组装本轮 Graph 输入请用 {@link buildTurnGraphMessages}（有 Checkpoint 历史则只追加 user）。
@@ -193,4 +197,5 @@ export declare function persistTurnTranscript(messageService: MessageService | u
     /** 无 assistant 文本消息时的兜底（invoke 的 graphResult.output） */
     fallbackAssistantContent?: string;
     assistantTokenCount?: number;
+    assistantDurationMs?: number | null;
 }): Promise<void>;

package/dist/session/persist-turn-messages.js

@@ -32,6 +32,7 @@ export async function persistAssistantTurnMessage(messageService, params) {
         reasoning: params.reasoningContent
             ? { reasoningContent: params.reasoningContent }
             : null,
+        durationMs: params.durationMs ?? null,
     });
 }
 /**
@@ -163,21 +164,32 @@ function isTextAssistant(message) {
         return false;
     return Boolean(message.content?.trim());
 }
+/** Session 只收成对的 tool 结果；code 节点没有 toolCallId，留在 Checkpoint。 */
+function isSessionToolMessage(message) {
+    if (message.role !== 'tool')
+        return false;
+    return Boolean(message.toolCallId?.trim());
+}
+function sessionGraphMessages(messages) {
+    return messages.filter((message) => message.role !== 'tool' || isSessionToolMessage(message));
+}
 /**
  * invoke 路径图增量。tool_calls / tool 行原样保留，纯文本助手收成一条放在末尾（与 stream 落库顺序一致）。
  * GraphMessage 无 nodeId：`finishPoints` 只含一个实际节点 id 时取最后一条纯文本，不是按 id 筛选（官方 DAG 的 `merge` 在末尾）。
  * 没有实际节点 id 或仅为 `__end__` 时把多段正文用空行拼进这一条。
+ * 无 toolCallId 的 tool 行不进 Session（code 节点结果）。
  */
 export function sessionVisibleGraphMessages(messages, finishPoints) {
-    const texts = messages.filter(isTextAssistant);
+    const eligible = sessionGraphMessages(messages);
+    const texts = eligible.filter(isTextAssistant);
     if (texts.length === 0)
-        return messages;
+        return eligible;
     const named = namedFinishPoints(finishPoints);
     const visible = named.length === 1 ? [texts.at(-1)] : texts;
     const keep = visible.at(-1);
     const joined = visible.map((message) => message.content).join('\n\n');
     const out = [];
-    for (const message of messages) {
+    for (const message of eligible) {
         if (isTextAssistant(message))
             continue;
         out.push(message);
@@ -344,8 +356,10 @@ export function extractHitlThreadId(output) {
     return undefined;
 }
 /**
- * Checkpoint thread_id 契约：HITL/多轮优先且应唯一等于 sessionId；
+ * Checkpoint 父图 thread_id：HITL/多轮优先且应唯一等于 sessionId；
  * 无 session 时退到 `task-${taskId}`。勿用 correlationId 作为 Checkpoint key。
+ * 子图槽位再经 `checkpointStorageThreadId`（`thread_id#父节点id`），
+ * 与 LangGraph put / 子图内 HITL interrupt 同一把钥匙。
  *
  * 注意：session 级 thread + 全量 Session 历史同时注入会导致 messages 双重 append。
  * 组装本轮 Graph 输入请用 {@link buildTurnGraphMessages}（有 Checkpoint 历史则只追加 user）。
@@ -436,14 +450,17 @@ export async function persistTurnTranscript(messageService, params) {
                 toolCalls: m.toolCalls,
                 tokenCount: isToolCallOnly ? 0 : params.assistantTokenCount,
                 reasoningContent: m.reasoningContent,
+                durationMs: isToolCallOnly ? null : params.assistantDurationMs,
             });
         }
         else if (m.role === 'tool') {
+            if (!m.toolCallId?.trim())
+                continue;
             await persistToolTurnMessage(messageService, {
                 sessionId: params.sessionId,
                 taskId: params.taskId,
                 correlationId: params.correlationId,
-                toolCallId: m.toolCallId ?? 'unknown',
+                toolCallId: m.toolCallId,
                 toolName: m.toolName ?? 'unknown',
                 content: m.content,
                 success: m.success !== false,
@@ -457,6 +474,7 @@ export async function persistTurnTranscript(messageService, params) {
             correlationId: params.correlationId,
             content: params.fallbackAssistantContent,
             tokenCount: params.assistantTokenCount,
+            durationMs: params.assistantDurationMs,
         });
     }
 }

package/dist/token/record-task-token-usage.d.ts

@@ -14,6 +14,29 @@ export declare function tokenUsageFromNodePayload(payload: unknown): TokenUsageR
 export declare function collectStreamNodeTokenUsage(event: Pick<RuntimeEvent, 'type' | 'payload'>, bucket: TokenUsageRecord[]): void;
 /** Graph `__llmTokenUsage` channel → Task.token_usage 行（invoke / fork 抛错后从 Checkpoint 读）。 */
 export declare function tokenUsagesFromLlmChannel(state: unknown): TokenUsageRecord[];
+/** 本轮 LLM 用量汇总，供 stats.token_usage 与 assistantTokenCount。 */
+export declare function summarizeTokenUsages(usages: Array<{
+    nodeId: string;
+    model: string;
+    inputTokens: number;
+    outputTokens: number;
+    costEstimate?: number;
+}>): {
+    total: {
+        input: number;
+        output: number;
+        total: number;
+    };
+    byNode: Record<string, {
+        input: number;
+        output: number;
+    }>;
+    byModel: Record<string, {
+        input: number;
+        output: number;
+    }>;
+    costEstimate: number;
+};
 export declare function recordTaskTokenUsages(taskService: TaskTokenUsageWriter, taskId: string, usages: Array<{
     nodeId: string;
     model: string;

package/dist/token/record-task-token-usage.js

@@ -53,6 +53,26 @@ export function tokenUsagesFromLlmChannel(state) {
     }
     return rows;
 }
+/** 本轮 LLM 用量汇总，供 stats.token_usage 与 assistantTokenCount。 */
+export function summarizeTokenUsages(usages) {
+    const byNode = {};
+    const byModel = {};
+    let input = 0;
+    let output = 0;
+    let costEstimate = 0;
+    for (const usage of usages) {
+        input += usage.inputTokens;
+        output += usage.outputTokens;
+        costEstimate += usage.costEstimate ?? 0;
+        const node = byNode[usage.nodeId] ?? (byNode[usage.nodeId] = { input: 0, output: 0 });
+        node.input += usage.inputTokens;
+        node.output += usage.outputTokens;
+        const model = byModel[usage.model] ?? (byModel[usage.model] = { input: 0, output: 0 });
+        model.input += usage.inputTokens;
+        model.output += usage.outputTokens;
+    }
+    return { total: { input, output, total: input + output }, byNode, byModel, costEstimate };
+}
 export async function recordTaskTokenUsages(taskService, taskId, usages) {
     for (const usage of usages) {
         try {

package/dist/tool/tool-confirmation.js

@@ -9,6 +9,7 @@
  * @see tasks.md Task 17
  */
 import { DEFAULT_HITL_TIMEOUT_MS, HitlPausedError, } from '../graph/hitl/hitl-handler.js';
+import { checkpointStorageThreadId } from '../graph/adapters/langgraph-checkpointer.js';
 import { resolveCheckpointThreadId } from '../session/persist-turn-messages.js';
 export class ToolConfirmationInterceptor {
     confirmationTools;
@@ -54,11 +55,13 @@ export class ToolConfirmationInterceptor {
                 errorCode: 'hitl_rejected',
             };
         }
-        const resolvedThreadId = threadId ??
-            resolveCheckpointThreadId({
-                sessionId: typeof state.sessionId === 'string' ? state.sessionId : undefined,
-                taskId,
-            });
+        const sessionId = typeof state.sessionId === 'string' && state.sessionId.trim()
+            ? state.sessionId.trim()
+            : undefined;
+        const parentThreadId = sessionId ??
+            threadId ??
+            resolveCheckpointThreadId({ taskId });
+        const resolvedThreadId = checkpointStorageThreadId(parentThreadId, typeof state.checkpointNs === 'string' ? state.checkpointNs : undefined);
         const question = `确认执行 ${request.toolName}？\n参数: ${JSON.stringify(request.arguments)}`;
         const options = ['confirm', 'reject'];
         const timeoutMs = DEFAULT_HITL_TIMEOUT_MS;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gt-fe/eap-runtime",
-  "version": "0.0.21",
+  "version": "0.0.22",
   "description": "Enterprise Agent Platform — Runtime核心：Graph执行引擎、Task生命周期、Session管理、Checkpoint持久化",
   "type": "module",
   "main": "dist/index.js",