@gt-fe/eap-runtime 0.0.20 → 0.0.22

package/dist/providers/routing-llm-provider.js

@@ -7,13 +7,26 @@
  * 3. 用返回的连接配置创建 HttpLLMProvider，发送请求
  *
  * 设计原则：
- * - Runtime 不存任何 LLM 凭证，全部从 Foundation 获取
+ * - 生产路径：Foundation `/models/select` 返回连接，本机 HTTP
+ * - select 失败或没有连接：走构造时注入的 fallback（dev 为 LocalLlmProvider / `resolveLocalLlmConnection`）
  * - 支持任意 OpenAI 兼容 API（OpenAI / DeepSeek / 本地 Ollama 等）
- * - 连接配置由 Foundation 统一管理，Runtime 无感知
  */
-import { mergeModelCallParams } from '@gt-fe/eap-contracts';
+import { AppError, ErrorCodes, mergeModelCallParams } from '@gt-fe/eap-contracts';
 import { EchoLLMProvider } from './echo-llm-provider.js';
 import { DEFAULT_LLM_HTTP_TIMEOUT_MS, HttpLLMProvider } from './http-llm-provider.js';
+export const MODELS_INVOKE_STREAM_MESSAGE = 'EAP_RUNTIME_MODELS_INVOKE 实验口只支持同步 invoke（POST /models/invoke），不支持 stream。关掉该开关，或改走 Foundation /models/select + 本机 HTTP。';
+export function isFoundationModelsInvokeEnabled(env = process.env) {
+    const raw = env.EAP_RUNTIME_MODELS_INVOKE;
+    return raw === '1' || raw === 'true';
+}
+/** `/models/invoke` 无真流式；开了该开关时 RoutingLLMProvider.stream 必须失败。 */
+export function assertFoundationModelsInvokeAllowsStream(env = process.env) {
+    if (!isFoundationModelsInvokeEnabled(env))
+        return;
+    throw new AppError(ErrorCodes.GRAPH_EXECUTION_ERROR, MODELS_INVOKE_STREAM_MESSAGE, {
+        exposeMessage: true,
+    });
+}
 export class RoutingLLMProvider {
     foundationUrl;
     fallback;
@@ -29,6 +42,7 @@ export class RoutingLLMProvider {
         return provider.invoke(messages, this.withCallDefaults(config, selected));
     }
     async *stream(messages, config) {
+        assertFoundationModelsInvokeAllowsStream();
         const selected = await this.selectModel(config);
         const provider = this.createProviderForModel(selected);
         yield* provider.stream(messages, this.withCallDefaults(config, selected));
@@ -38,7 +52,7 @@ export class RoutingLLMProvider {
         return {
             ...config,
             ...merged,
-            model: selected.name,
+            model: selected.name || config.model,
             maxContextTokens: selected.maxContextTokens ?? config.maxContextTokens,
         };
     }
@@ -70,10 +84,15 @@ export class RoutingLLMProvider {
         catch {
             // fallback to local
         }
-        return { name: config.model ?? 'echo', provider: 'local', connection: null };
+        return { name: config.model ?? '', provider: 'local', connection: null };
+    }
+    useFoundationInvoke() {
+        return isFoundationModelsInvokeEnabled();
     }
     createProviderForModel(selected) {
-        // 优先使用 Foundation 返回的连接配置
+        if (this.useFoundationInvoke()) {
+            return new FoundationModelsInvokeProvider(this.foundationUrl, selected.name, this.turnLog);
+        }
         if (selected.connection) {
             return new HttpLLMProvider({
                 apiKey: selected.connection.apiKey ?? '',
@@ -83,16 +102,52 @@ export class RoutingLLMProvider {
                 turnLog: this.turnLog,
             });
         }
-        // Fallback: 兼容旧逻辑（直接读 env）—— 开发环境快速启动用
-        if (process.env.OPENAI_API_KEY && selected.provider === 'openai') {
-            return new HttpLLMProvider({
-                apiKey: process.env.OPENAI_API_KEY,
-                baseUrl: process.env.OPENAI_BASE_URL,
-                defaultModel: selected.name,
-                timeoutMs: DEFAULT_LLM_HTTP_TIMEOUT_MS,
-                turnLog: this.turnLog,
-            });
-        }
+        // Foundation 未返回连接：与 LocalLlmProvider 同一把钥匙（由调用方注入 fallback）。
         return this.fallback;
     }
 }
+/**
+ * Foundation `/models/invoke` 实验口。默认关。
+ * 只实现同步 invoke；stream 拒绝（无真流式、无鉴权头）。
+ * Foundation 未提供该接口时不要开 EAP_RUNTIME_MODELS_INVOKE。
+ */
+class FoundationModelsInvokeProvider {
+    foundationUrl;
+    model;
+    turnLog;
+    constructor(foundationUrl, model, turnLog) {
+        this.foundationUrl = foundationUrl;
+        this.model = model;
+        this.turnLog = turnLog;
+    }
+    async invoke(messages, config) {
+        const res = await fetch(`${this.foundationUrl}/api/v1/models/invoke`, {
+            method: 'POST',
+            headers: { 'Content-Type': 'application/json' },
+            body: JSON.stringify({
+                model: config.model ?? this.model,
+                messages,
+                temperature: config.temperature,
+                maxTokens: config.maxTokens,
+            }),
+        });
+        if (!res.ok) {
+            throw new Error(`Foundation /models/invoke HTTP ${res.status}`);
+        }
+        const body = (await res.json());
+        const content = body.data?.content ?? body.content ?? '';
+        void this.turnLog;
+        return {
+            content,
+            inputTokens: body.data?.inputTokens ?? 0,
+            outputTokens: body.data?.outputTokens ?? 0,
+            model: body.data?.model ?? config.model ?? this.model,
+        };
+    }
+    async *stream(_messages, _config) {
+        assertFoundationModelsInvokeAllowsStream();
+        throw new AppError(ErrorCodes.GRAPH_EXECUTION_ERROR, MODELS_INVOKE_STREAM_MESSAGE, {
+            exposeMessage: true,
+        });
+    }
+}

package/dist/registry/deployment-snapshot.js

@@ -116,7 +116,7 @@ async function findDigestMatch(items, expectedRaw) {
     }
     return undefined;
 }
-async function warnIfDigestMismatch(expectedRaw, actualHex, label, details) {
+async function warnIfDigestMismatch(expectedRaw, actualHex, label, details, fail = false) {
     const expected = normalizeDigest(expectedRaw);
     if (!expected)
         return;
@@ -129,6 +129,13 @@ async function warnIfDigestMismatch(expectedRaw, actualHex, label, details) {
         actual,
         ...details,
     });
+    if (fail) {
+        throw preloadPackageCorrupt(`预装校验和不匹配: ${label}`, {
+            expected,
+            actual,
+            ...details,
+        });
+    }
 }
 async function pickLatestVersionDir(parent) {
     if (!(await pathExists(parent)))
@@ -373,6 +380,13 @@ export class FileSystemDeploymentSnapshot {
                         actual: fileDigest,
                         path: pkgFile,
                     });
+                    if (this.flags.requireChecksum || isRequireChecksum()) {
+                        throw preloadPackageCorrupt('snapshot.sha256 与 .package.sha256 不一致', {
+                            expected: expectedPkg,
+                            actual: fileDigest,
+                            path: pkgFile,
+                        });
+                    }
                 }
             }
         }
@@ -394,7 +408,7 @@ export class FileSystemDeploymentSnapshot {
                 throw preloadPackageCorrupt(`artifacts 指向的文件不存在: ${rel}`, { path: full });
             }
             const actual = await sha256OfFile(full);
-            await warnIfDigestMismatch(expected, actual, `artifact ${rel}`, { path: full });
+            await warnIfDigestMismatch(expected, actual, `artifact ${rel}`, { path: full }, this.flags.requireChecksum || isRequireChecksum());
         }
     }
     async resolveAgentDir(agentCode, version) {
@@ -499,7 +513,7 @@ export class FileSystemDeploymentSnapshot {
             if (ref.sha256) {
                 const matched = await findDigestMatch(existingManifests, ref.sha256);
                 if (!matched) {
-                    await warnIfDigestMismatch(ref.sha256, await sha256OfFile(existingManifests[0].path), `tool ${ref.name}`, { path: existingManifests[0].path });
+                    await warnIfDigestMismatch(ref.sha256, await sha256OfFile(existingManifests[0].path), `tool ${ref.name}`, { path: existingManifests[0].path }, this.flags.requireChecksum || isRequireChecksum());
                 }
                 chosen = matched ?? existingManifests[0];
             }
@@ -576,7 +590,7 @@ export class FileSystemDeploymentSnapshot {
                 !isAuthorOverlay &&
                 manifestPath.endsWith('.json')) {
                 const actual = await sha256OfFile(manifestPath);
-                await warnIfDigestMismatch(ref.sha256, actual, `skill ${ref.skillId}`, { path: manifestPath });
+                await warnIfDigestMismatch(ref.sha256, actual, `skill ${ref.skillId}`, { path: manifestPath }, this.flags.requireChecksum || isRequireChecksum());
             }
             const skillId = String(ref.skillId);
             return {

package/dist/registry/graph-loader.d.ts

@@ -108,6 +108,8 @@ export declare class GraphLoader {
     definitionCacheSize(): number;
     private cloneCached;
     loadGraph(agentCode: string, options?: LoadAgentOptions): Promise<GraphDSL>;
+    /** Agent `models.default`；无清单或未声明则 undefined。 */
+    resolveDefaultModel(agentCode: string, version?: string): Promise<string | undefined>;
     private loadGraphFromSource;
     /**
      * Agent 基线系统提示词（manifest.prompts.system）。

package/dist/registry/graph-loader.js

@@ -144,6 +144,26 @@ const DEFAULT_GRAPH_DSL = {
     finishPoints: ['__end__'],
     config: { timeout: 60000 },
 };
+/** 源图有非空字符串则保留；缺省或空数组为 `['__end__']`。预装 graph.json 与清单内嵌图同一收口。 */
+function finishPointsFromGraphDocument(graph) {
+    const raw = graph.finishPoints;
+    if (!Array.isArray(raw))
+        return ['__end__'];
+    const ids = raw
+        .filter((id) => typeof id === 'string')
+        .map((id) => id.trim())
+        .filter((id) => id.length > 0);
+    return ids.length > 0 ? ids : ['__end__'];
+}
+const DEFAULT_LOADED_GRAPH_CONFIG = { timeout: 60000 };
+/** 默认 timeout 60000；源图 `config` 覆盖。缺 `maxIterations` 时由引擎用 50。 */
+function configFromGraphDocument(graph) {
+    const raw = graph.config;
+    if (!raw || typeof raw !== 'object' || Array.isArray(raw)) {
+        return { ...DEFAULT_LOADED_GRAPH_CONFIG };
+    }
+    return { ...DEFAULT_LOADED_GRAPH_CONFIG, ...raw };
+}
 function yamlGraphToDsl(agentCode, graph) {
     const nodes = graph.nodes ?? [];
     const edges = graph.edges ?? [];
@@ -155,8 +175,8 @@ function yamlGraphToDsl(agentCode, graph) {
         edges,
         conditionalEdges: graph.conditionalEdges,
         entryPoint: graph.entryPoint ?? nodes[0]?.id ?? 'llm',
-        finishPoints: ['__end__'],
-        config: { timeout: 60000 },
+        finishPoints: finishPointsFromGraphDocument(graph),
+        config: configFromGraphDocument(graph),
     };
 }
 function collectToolNames(agent, graphDsl) {
@@ -620,6 +640,23 @@ export class GraphLoader {
         }
         return this.cloneCached(graph);
     }
+    /** Agent `models.default`；无清单或未声明则 undefined。 */
+    async resolveDefaultModel(agentCode, version) {
+        const requested = this.resolveVersion({ agentVersion: version });
+        if (this.deploymentSnapshot) {
+            try {
+                const snap = await this.deploymentSnapshot.loadAgentSnapshot(agentCode, requested);
+                const model = resolveAgentDefaultModel((snap?.manifest ?? {}));
+                const trimmed = model?.trim();
+                if (trimmed && trimmed.toLowerCase() !== 'default')
+                    return trimmed;
+            }
+            catch {
+                return undefined;
+            }
+        }
+        return undefined;
+    }
     async loadGraphFromSource(agentCode, version) {
         if (this.deploymentSnapshot) {
             const snap = await this.deploymentSnapshot.loadAgentSnapshot(agentCode, version);

package/dist/registry/preload-flags.d.ts

@@ -5,7 +5,7 @@
  * |------|------|------|
  * | EAP_PRELOAD_REQUIRE_READY | 关 | 强制校验 .ready |
  * | EAP_PRELOAD_VALIDATE_SKILL_IO | **开** | 执行子图前校验 skill inputDef.required；`0`/`false` 关闭 |
- * | EAP_PRELOAD_REQUIRE_CHECKSUM | 关 | snapshot 必须带 sha256 或 artifacts[]（字段存在性）。加载期摘要不匹配为 warn |
+ * | EAP_PRELOAD_REQUIRE_CHECKSUM | 关 | snapshot 必须带 sha256 或 artifacts[]。开启时摘要不匹配抛 RUNTIME_PRELOAD_PACKAGE_CORRUPT；未开仍 warn |
  * | EAP_PRELOAD_DEFINITION_CACHE | 有预装时开 | 进程内定义缓存；`0`/`false` 关闭 |
  * | EAP_SKILL_HOT_RELOAD | **dev 开 / server 关** | 已装配 Skill 的目录 mtime 变化则丢缓存；未写入装配名单的目录不进清单 |
  * | EAP_LOCAL_SKILLS_DIR | 关 | 开发态作者源 Skill 根；查找优先于 preinstall，低于 users/ |

package/dist/registry/preload-flags.js

@@ -5,7 +5,7 @@
  * |------|------|------|
  * | EAP_PRELOAD_REQUIRE_READY | 关 | 强制校验 .ready |
  * | EAP_PRELOAD_VALIDATE_SKILL_IO | **开** | 执行子图前校验 skill inputDef.required；`0`/`false` 关闭 |
- * | EAP_PRELOAD_REQUIRE_CHECKSUM | 关 | snapshot 必须带 sha256 或 artifacts[]（字段存在性）。加载期摘要不匹配为 warn |
+ * | EAP_PRELOAD_REQUIRE_CHECKSUM | 关 | snapshot 必须带 sha256 或 artifacts[]。开启时摘要不匹配抛 RUNTIME_PRELOAD_PACKAGE_CORRUPT；未开仍 warn |
  * | EAP_PRELOAD_DEFINITION_CACHE | 有预装时开 | 进程内定义缓存；`0`/`false` 关闭 |
  * | EAP_SKILL_HOT_RELOAD | **dev 开 / server 关** | 已装配 Skill 的目录 mtime 变化则丢缓存；未写入装配名单的目录不进清单 |
  * | EAP_LOCAL_SKILLS_DIR | 关 | 开发态作者源 Skill 根；查找优先于 preinstall，低于 users/ |

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

@@ -16,6 +16,7 @@ import type { SessionContextService } from '../session/session-context.service.j
 import type { MessageService } from '../session/message.service.js';
 import type { SessionService, SessionRecord, MessageRecord } from '../session/session.service.js';
 import { type ILlmTurnLog, type LlmTurnIndexEntry, type LlmTurnLogKind, type LlmTurnRecord } from '../debug/llm-turn-log.js';
+import { type SessionCompressionOptions } from '../session/apply-session-compression.js';
 import type { SSEEvent } from '../sse/contracts/sse-emitter.contract.js';
 import type { TaskService } from '../task/task.service.js';
 import type { ExecuteParams, ExecuteResult, ForkParams, ForkOverrides, ForkResult, HitlResponse, ResumeHitlResult, SeedSessionMessagesParams, SeedSessionMessagesResult } from './types.js';
@@ -46,6 +47,10 @@ export interface AgentRuntimeDeps {
      */
     traceCollector?: ITraceCollector;
     recordForkRelation?: RecordForkRelation;
+    /** 历史压缩；默认关。见 EAP_RUNTIME_COMPRESSION */
+    sessionCompression?: SessionCompressionOptions;
+    /** 长期记忆注入；默认关。见 EAP_RUNTIME_MEMORY */
+    sessionMemory?: import('../session/apply-session-memory.js').SessionMemoryOptions;
 }
 /** 一次对话详情：消息 + 每步 LLM 入参/响应。File / 内存 / Postgres 同一套结构。 */
 export interface ConversationDetail {

package/dist/runtime/agent-runtime.js

@@ -7,17 +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, extractStreamTextDelta, messageRecordsToGraphMessages, persistHitlWaitingPlaceholder, persistTurnTranscript, persistUserTurnMessage, } from '../session/persist-turn-messages.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, 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, } 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',
@@ -208,6 +211,8 @@ export class AgentRuntime {
             executionContext: ctx,
             userId,
             traceId: resolveTraceId(params),
+            ragScope: params.ragScope,
+            ragSources: params.ragSources,
         }, userId);
         return response.data;
     }
@@ -274,7 +279,34 @@ export class AgentRuntime {
             ? (await this.deps.graphLoader.resolveAgentDir(agentCode, params.agentVersion)) ?? undefined
             : undefined;
         const agentSystemPrompt = await resolveAgentSystemPromptFromLoader(this.deps.graphLoader, agentCode, params.agentVersion);
-        const history = messageRecordsToGraphMessages(sessionContext.messages);
+        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: compressionOptions,
+        })) {
+            if (step.kind === 'history') {
+                history = step.messages;
+                continue;
+            }
+            yield this.toSSE(step.kind === 'started' ? 'session.compression.started' : 'session.compression.completed', correlationId, task.id, { ...step.payload });
+        }
+        history = await applySessionMemory({
+            history,
+            userId,
+            agentCode,
+            query: typeof params.input === 'string' ? params.input : undefined,
+            options: this.deps.sessionMemory,
+        });
+        const tokenDetail = resolveTokenDetail(params.config?.tokenDetail);
         const messages = await buildTurnGraphMessages({
             history,
             input: params.input,
@@ -282,7 +314,7 @@ export class AgentRuntime {
             threadId: sessionId,
         });
         const startTime = Date.now();
-        let assistantText = '';
+        const assistantBuckets = createStreamAssistantBuckets();
         let toolCallReasoningContent;
         let assistantReasoningContent;
         const pendingToolCalls = [];
@@ -300,6 +332,8 @@ export class AgentRuntime {
                 sessionId,
                 executionContext: ctx,
                 forceToolServer: params.config?.forceToolServer,
+                ragScope: params.ragScope,
+                ragSources: params.ragSources,
                 loadedTools,
                 loadedSkills,
                 skillDefinitions,
@@ -323,35 +357,20 @@ export class AgentRuntime {
                         toolCallReasoningContent = p.reasoningContent;
                     }
                 }
-                if (runtimeEvent.type === 'tool.result') {
-                    const p = runtimeEvent.payload;
-                    if (p.toolCallId && p.toolName) {
-                        toolResults.push({
-                            toolCallId: p.toolCallId,
-                            toolName: p.toolName,
-                            content: p.content,
-                            success: p.success !== false,
-                        });
-                    }
-                }
+                // 终态 tool.result / tool.error 回填历史；willRetry 的 error 不进 transcript
+                appendTerminalToolResult(runtimeEvent.type, runtimeEvent.payload, toolResults);
                 const streamedReasoning = extractStreamReasoning(runtimeEvent.type, runtimeEvent.payload);
                 if (streamedReasoning)
                     assistantReasoningContent = streamedReasoning;
-                const delta = extractStreamTextDelta(runtimeEvent.type, runtimeEvent.payload);
-                if (delta) {
-                    if (runtimeEvent.type === 'response.final') {
-                        if (delta.trim())
-                            assistantText = delta;
-                    }
-                    else {
-                        assistantText += delta;
-                    }
-                }
+                applyStreamTextDelta(assistantBuckets, runtimeEvent.type, runtimeEvent.payload);
                 collectStreamNodeTokenUsage(runtimeEvent, nodeTokenUsages);
-                yield this.runtimeEventToSSE(runtimeEvent, correlationId, task.id);
+                if (shouldEmitTokenSse(tokenDetail, runtimeEvent.type)) {
+                    yield this.runtimeEventToSSE(runtimeEvent, correlationId, task.id);
+                }
             }
             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,
@@ -359,11 +378,18 @@ export class AgentRuntime {
                 messages: buildTurnTranscriptFromStreamEvents({
                     pendingToolCalls,
                     toolResults,
-                    assistantText,
+                    assistantTexts: streamAssistantTexts(assistantBuckets),
+                    assistantText: visibleStreamAssistantText(assistantBuckets),
+                    finishPoints: graph.finishPoints,
                     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,
@@ -390,6 +416,7 @@ export class AgentRuntime {
                 return;
             }
             const message = err instanceof Error ? err.message : 'Unknown error';
+            await recordTaskTokenUsages(this.deps.taskService, task.id, nodeTokenUsages);
             await this.deps.taskService.updateStatus(task.id, 'failed', message);
             yield this.toSSE('task.error', correlationId, task.id, {
                 code: 'runtime.graph.execution_error',
@@ -409,7 +436,7 @@ export class AgentRuntime {
         let output = '';
         let waiting;
         let sawWait = false;
-        let assistantText = '';
+        const assistantBuckets = createStreamAssistantBuckets();
         for await (const event of this.streamResumeHitl(taskId, response)) {
             if (event.type === 'hitl.waited') {
                 sawWait = true;
@@ -421,21 +448,12 @@ export class AgentRuntime {
                 };
                 output = waiting.question;
             }
-            const delta = extractStreamTextDelta(event.type, event.data);
-            if (delta) {
-                if (event.type === 'response.final') {
-                    if (delta.trim())
-                        assistantText = delta;
-                }
-                else {
-                    assistantText += delta;
-                }
-            }
+            applyStreamTextDelta(assistantBuckets, event.type, event.data);
         }
         if (sawWait) {
             return { taskId, status: 'waiting_hitl', output, hitl: waiting };
         }
-        return { taskId, status: 'success', output: assistantText || output };
+        return { taskId, status: 'success', output: visibleStreamAssistantText(assistantBuckets) || output };
     }
     /**
      * HITL 恢复的流式入口：新开一条 SSE（第一条 stream 在 `hitl.waited` 后已结束）。
@@ -465,6 +483,7 @@ export class AgentRuntime {
         if (!claimed) {
             throw new AppError(ErrorCodes.TASK_HITL_INVALID_STATE, `Task ${taskId} is no longer waiting for HITL`, { exposeMessage: true });
         }
+        const tokenDetail = resolveTokenDetail(response.tokenDetail);
         yield this.toSSE('hitl.resumed', correlationId, taskId, {
             choice: response.choice,
             userId: response.userId,
@@ -516,7 +535,7 @@ export class AgentRuntime {
             traceCollector: this.deps.traceCollector,
         };
         const startTime = Date.now();
-        let assistantText = '';
+        const assistantBuckets = createStreamAssistantBuckets();
         let toolCallReasoningContent;
         let assistantReasoningContent;
         const pendingToolCalls = [];
@@ -537,35 +556,19 @@ export class AgentRuntime {
                         toolCallReasoningContent = p.reasoningContent;
                     }
                 }
-                if (runtimeEvent.type === 'tool.result' || runtimeEvent.type === 'tool.error') {
-                    const p = runtimeEvent.payload;
-                    if (p.toolCallId && p.toolName) {
-                        toolResults.push({
-                            toolCallId: p.toolCallId,
-                            toolName: p.toolName,
-                            content: p.content ?? p.error,
-                            success: runtimeEvent.type === 'tool.result' && p.success !== false,
-                        });
-                    }
-                }
+                appendTerminalToolResult(runtimeEvent.type, runtimeEvent.payload, toolResults);
                 const streamedReasoning = extractStreamReasoning(runtimeEvent.type, runtimeEvent.payload);
                 if (streamedReasoning)
                     assistantReasoningContent = streamedReasoning;
-                const delta = extractStreamTextDelta(runtimeEvent.type, runtimeEvent.payload);
-                if (delta) {
-                    if (runtimeEvent.type === 'response.final') {
-                        if (delta.trim())
-                            assistantText = delta;
-                    }
-                    else {
-                        assistantText += delta;
-                    }
-                }
+                applyStreamTextDelta(assistantBuckets, runtimeEvent.type, runtimeEvent.payload);
                 collectStreamNodeTokenUsage(runtimeEvent, nodeTokenUsages);
-                yield this.runtimeEventToSSE(runtimeEvent, correlationId, taskId);
+                if (shouldEmitTokenSse(tokenDetail, runtimeEvent.type)) {
+                    yield this.runtimeEventToSSE(runtimeEvent, correlationId, taskId);
+                }
             }
             await recordTaskTokenUsages(this.deps.taskService, taskId, nodeTokenUsages);
-            await this.deps.taskService.updateStatus(taskId, 'success', assistantText || undefined);
+            await this.deps.taskService.updateStatus(taskId, 'success', visibleStreamAssistantText(assistantBuckets) || undefined);
+            const tokenSummary = summarizeTokenUsages(nodeTokenUsages);
             if (sessionId) {
                 await persistTurnTranscript(this.deps.messageService, {
                     sessionId,
@@ -574,12 +577,19 @@ export class AgentRuntime {
                     messages: buildTurnTranscriptFromStreamEvents({
                         pendingToolCalls,
                         toolResults,
-                        assistantText,
+                        assistantTexts: streamAssistantTexts(assistantBuckets),
+                        assistantText: visibleStreamAssistantText(assistantBuckets),
+                        finishPoints: graph.finishPoints,
                         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,
@@ -609,6 +619,7 @@ export class AgentRuntime {
             }
             const appErr = err instanceof AppError ? err : AppError.fromUnknown(err);
             const failMsg = appErr.exposeMessage ? appErr.message : '系统内部错误';
+            await recordTaskTokenUsages(this.deps.taskService, taskId, nodeTokenUsages);
             await this.deps.taskService.updateStatus(taskId, 'failed', failMsg);
             yield this.toSSE('task.error', correlationId, taskId, {
                 code: appErr.code,
@@ -671,8 +682,14 @@ export class AgentRuntime {
             ? forkedCheckpoint.state.agentSystemPrompt
             : undefined);
         try {
+            const forkState = { ...forkedCheckpoint.state };
+            delete forkState.toolSpanId;
+            delete forkState.hitlToolCallId;
             const result = await this.deps.graphEngine.invoke(graph, {
-                ...forkedCheckpoint.state,
+                messages: Array.isArray(forkState.messages)
+                    ? forkState.messages
+                    : [],
+                ...forkState,
                 agentVersion,
                 agentDir,
                 agentSystemPrompt,
@@ -729,6 +746,20 @@ export class AgentRuntime {
         }
         catch (err) {
             const message = err instanceof Error ? err.message : 'Unknown error';
+            const sessionId = typeof forkedCheckpoint.state.sessionId === 'string'
+                ? forkedCheckpoint.state.sessionId
+                : undefined;
+            const threadId = resolveCheckpointThreadId({
+                sessionId,
+                taskId: task.id,
+            });
+            const ck = await this.deps.checkpointSaver.load(threadId);
+            const baseline = tokenUsagesFromLlmChannel(forkedCheckpoint.state);
+            const latest = tokenUsagesFromLlmChannel(ck?.state ?? forkedCheckpoint.state);
+            const usages = latest.length >= baseline.length ? latest.slice(baseline.length) : latest;
+            await recordTaskTokenUsages(this.deps.taskService, task.id, usages);
+            const input_tokens = usages.reduce((sum, row) => sum + row.inputTokens, 0);
+            const output_tokens = usages.reduce((sum, row) => sum + row.outputTokens, 0);
             await this.deps.taskService.updateStatus(task.id, 'failed', message);
             return {
                 newTraceId: traceId,
@@ -737,7 +768,11 @@ export class AgentRuntime {
                 checkpointId,
                 status: 'failed',
                 output: message,
-                tokenUsage: { input_tokens: 0, output_tokens: 0, total_tokens: 0 },
+                tokenUsage: {
+                    input_tokens,
+                    output_tokens,
+                    total_tokens: input_tokens + output_tokens,
+                },
             };
         }
     }
@@ -758,16 +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 &&