@gt-fe/eap-runtime 0.0.23 → 0.0.25

package/dist/runtime/agent-runtime.js

@@ -7,12 +7,15 @@
  * @see 04-runtime-core/design.md §1.5 / §1.7
  */
 import { randomUUID } from 'node:crypto';
+import { DEFAULT_GRAPH_TIMEOUT_MS } from '../common/types.js';
+import { resolveTaskTimeoutMs } from '../task/resolve-task-timeout.js';
+import { runWithExecutionTrace } from '../graph/execution-trace-context.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 { buildHitlWaitingTaskOutput, buildTurnTranscriptFromStreamEvents, loadModelFacingHistory, turnGraphMessagesAfterCompression, 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';
@@ -37,7 +40,7 @@ const DEFAULT_GRAPH_DSL = {
         { from: '__start__', to: 'llm' },
         { from: 'llm', to: '__end__' },
     ],
-    config: { timeout: 60000 },
+    config: { timeout: DEFAULT_GRAPH_TIMEOUT_MS },
 };
 /**
  * AgentRuntime — 管理 Graph 执行的 SDK 门面
@@ -214,6 +217,14 @@ export class AgentRuntime {
             ragScope: params.ragScope,
             ragSources: params.ragSources,
         }, userId);
+        if (response.code !== 'ok') {
+            const output = typeof response.data?.output === 'string' && response.data.output
+                ? response.data.output
+                : response.code;
+            throw new AppError(response.code, output, {
+                exposeMessage: true,
+            });
+        }
         return response.data;
     }
     /**
@@ -241,6 +252,11 @@ export class AgentRuntime {
                 userId,
             });
         }
+        const loadOpts = { agentVersion: params.agentVersion };
+        const graphDsl = this.deps.graphLoader
+            ? await this.deps.graphLoader.loadGraph(agentCode, loadOpts)
+            : DEFAULT_GRAPH_DSL;
+        const timeout = resolveTaskTimeoutMs(params.config?.timeout, graphDsl.config?.timeout);
         const task = await this.deps.taskService.create({
             id: params.taskId,
             agentId: agentCode,
@@ -248,9 +264,12 @@ export class AgentRuntime {
             userId,
             correlationId,
             input: params.input,
-            timeout: params.config?.timeout ?? 60_000,
+            timeout,
         });
         await this.deps.taskService.updateStatus(task.id, 'running');
+        const abortSignal = this.deps.timeoutManager?.startTimer(task.id, timeout, async () => {
+            await this.deps.taskService.tryUpdateStatus(task.id, 'running', 'timeout', 'Task execution timed out');
+        });
         await persistUserTurnMessage(this.deps.messageService, {
             sessionId,
             taskId: task.id,
@@ -258,10 +277,6 @@ export class AgentRuntime {
             content: params.input,
         });
         yield this.toSSE('task.started', correlationId, task.id, { taskId: task.id, sessionId });
-        const loadOpts = { agentVersion: params.agentVersion };
-        const graphDsl = this.deps.graphLoader
-            ? await this.deps.graphLoader.loadGraph(agentCode, loadOpts)
-            : DEFAULT_GRAPH_DSL;
         const graph = await this.deps.graphEngine.buildGraph(graphDsl);
         const loadedToolsBase = await this.resolveLoadedTools(agentCode, params.agentVersion);
         const loadedSkills = this.deps.graphLoader
@@ -279,7 +294,13 @@ export class AgentRuntime {
             ? (await this.deps.graphLoader.resolveAgentDir(agentCode, params.agentVersion)) ?? undefined
             : undefined;
         const agentSystemPrompt = await resolveAgentSystemPromptFromLoader(this.deps.graphLoader, agentCode, params.agentVersion);
-        let history = messageRecordsToGraphMessages(sessionContext.messages);
+        const facing = await loadModelFacingHistory({
+            sessionHistory: messageRecordsToGraphMessages(sessionContext.messages),
+            checkpointSaver: this.deps.checkpointSaver,
+            threadId: sessionId,
+        });
+        let history = facing.messages;
+        let compressed = false;
         const compressionOptions = await withCompressionSummaryModel({
             ...this.deps.sessionCompression,
             turnLog: this.deps.sessionCompression?.turnLog ?? this.deps.turnLog,
@@ -297,6 +318,9 @@ export class AgentRuntime {
                 history = step.messages;
                 continue;
             }
+            if (step.kind === 'completed') {
+                compressed = step.payload.success === true;
+            }
             yield this.toSSE(step.kind === 'started' ? 'session.compression.started' : 'session.compression.completed', correlationId, task.id, { ...step.payload });
         }
         history = await applySessionMemory({
@@ -307,11 +331,11 @@ export class AgentRuntime {
             options: this.deps.sessionMemory,
         });
         const tokenDetail = resolveTokenDetail(params.config?.tokenDetail);
-        const messages = await buildTurnGraphMessages({
+        const { messages, replaceCheckpointMessages } = turnGraphMessagesAfterCompression({
             history,
             input: params.input,
-            checkpointSaver: this.deps.checkpointSaver,
-            threadId: sessionId,
+            fromCheckpoint: facing.fromCheckpoint,
+            compressed,
         });
         const startTime = Date.now();
         const assistantBuckets = createStreamAssistantBuckets();
@@ -321,8 +345,9 @@ export class AgentRuntime {
         const toolResults = [];
         const nodeTokenUsages = [];
         try {
-            for await (const runtimeEvent of this.deps.graphEngine.stream(graph, {
+            for await (const runtimeEvent of runWithExecutionTrace({ abortSignal }, () => this.deps.graphEngine.stream(graph, {
                 messages,
+                replaceCheckpointMessages,
                 userId,
                 delegationToken: ctx.delegationToken,
                 agentCode,
@@ -343,7 +368,9 @@ export class AgentRuntime {
                 skillInput: resolveSkillInputFromInvokeInput(params.input),
                 traceId: resolveTraceId(params),
                 traceCollector: this.deps.traceCollector,
-            })) {
+                abortSignal,
+                abortDeadlineMs: Date.now() + timeout,
+            }))) {
                 if (runtimeEvent.type === 'tool.called') {
                     const p = runtimeEvent.payload;
                     if (p.toolCallId && p.toolName) {
@@ -415,6 +442,17 @@ export class AgentRuntime {
                 yield this.toSSE('hitl.waited', correlationId, task.id, buildHitlWaitedPayload(hitlPaused));
                 return;
             }
+            const timedOut = err instanceof Error && /timed out/i.test(err.message);
+            if (timedOut) {
+                await recordTaskTokenUsages(this.deps.taskService, task.id, nodeTokenUsages);
+                await this.deps.taskService.tryUpdateStatus(task.id, 'running', 'timeout');
+                yield this.toSSE('task.error', correlationId, task.id, {
+                    code: ErrorCodes.TASK_TIMEOUT,
+                    message: 'Task execution timed out',
+                    recoverable: false,
+                });
+                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);
@@ -425,6 +463,9 @@ export class AgentRuntime {
             });
             throw err;
         }
+        finally {
+            this.deps.timeoutManager?.clearTimer(task.id);
+        }
     }
     /**
      * HITL 用户响应后继续执行（同步结果）。
@@ -483,6 +524,9 @@ export class AgentRuntime {
         if (!claimed) {
             throw new AppError(ErrorCodes.TASK_HITL_INVALID_STATE, `Task ${taskId} is no longer waiting for HITL`, { exposeMessage: true });
         }
+        const abortSignal = this.deps.timeoutManager?.startTimer(taskId, resolveTaskTimeoutMs(task.timeout), async () => {
+            await this.deps.taskService.tryUpdateStatus(taskId, 'running', 'timeout', 'Task execution timed out');
+        });
         const tokenDetail = resolveTokenDetail(response.tokenDetail);
         yield this.toSSE('hitl.resumed', correlationId, taskId, {
             choice: response.choice,
@@ -507,6 +551,8 @@ export class AgentRuntime {
             : await this.resolveLoadedSkills(agentCode, response.userId, agentVersion);
         const continueState = {
             ...resumedState,
+            abortSignal,
+            abortDeadlineMs: Date.now() + resolveTaskTimeoutMs(task.timeout),
             messages: resumedState.messages ?? [],
             replaceCheckpointMessages: true,
             sessionId: resumedState.sessionId ?? sessionId,
@@ -542,7 +588,7 @@ export class AgentRuntime {
         const toolResults = [];
         const nodeTokenUsages = [];
         try {
-            for await (const runtimeEvent of this.deps.graphEngine.stream(graph, continueState)) {
+            for await (const runtimeEvent of runWithExecutionTrace({ abortSignal, abortDeadlineMs: continueState.abortDeadlineMs }, () => this.deps.graphEngine.stream(graph, continueState))) {
                 if (runtimeEvent.type === 'tool.called') {
                     const p = runtimeEvent.payload;
                     if (p.toolCallId && p.toolName) {
@@ -617,6 +663,17 @@ export class AgentRuntime {
                 yield this.toSSE('hitl.waited', correlationId, taskId, buildHitlWaitedPayload(hitlPaused));
                 return;
             }
+            const timedOut = err instanceof Error && /timed out/i.test(err.message);
+            if (timedOut) {
+                await recordTaskTokenUsages(this.deps.taskService, taskId, nodeTokenUsages);
+                await this.deps.taskService.tryUpdateStatus(taskId, 'running', 'timeout');
+                yield this.toSSE('task.error', correlationId, taskId, {
+                    code: ErrorCodes.TASK_TIMEOUT,
+                    message: 'Task execution timed out',
+                    recoverable: false,
+                });
+                return;
+            }
             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);
@@ -628,6 +685,9 @@ export class AgentRuntime {
             });
             throw err instanceof AppError ? err : appErr;
         }
+        finally {
+            this.deps.timeoutManager?.clearTimer(taskId);
+        }
     }
     /**
      * 从 Checkpoint Fork 并真实重新执行。
@@ -658,6 +718,12 @@ export class AgentRuntime {
         const newThreadId = randomUUID();
         const forkedCheckpoint = await this.deps.checkpointSaver.fork(checkpointId, newThreadId, overrides);
         const correlationId = overrides.executionContext?.correlationId ?? randomUUID();
+        const agentVersion = overrides.agentVersion ??
+            (typeof source.state.agentVersion === 'string' ? source.state.agentVersion : undefined);
+        const graphDsl = this.deps.graphLoader && typeof this.deps.graphLoader.loadGraph === 'function'
+            ? await this.deps.graphLoader.loadGraph(agentCode, { agentVersion })
+            : DEFAULT_GRAPH_DSL;
+        const timeout = resolveTaskTimeoutMs(undefined, graphDsl.config?.timeout);
         const task = await this.deps.taskService.create({
             id: taskId,
             agentId: agentCode,
@@ -665,12 +731,13 @@ export class AgentRuntime {
             userId,
             correlationId,
             input: { forkFrom: checkpointId, overrides },
-            timeout: 600_000,
+            timeout,
         });
         await this.deps.taskService.updateStatus(task.id, 'running');
-        const agentVersion = overrides.agentVersion ??
-            (typeof source.state.agentVersion === 'string' ? source.state.agentVersion : undefined);
-        const graph = await this.rebuildExecutableGraph(agentCode, agentVersion);
+        const abortSignal = this.deps.timeoutManager?.startTimer(task.id, timeout, async () => {
+            await this.deps.taskService.tryUpdateStatus(task.id, 'running', 'timeout', 'Task execution timed out');
+        });
+        const graph = await this.deps.graphEngine.buildGraph(graphDsl);
         const loadedToolsBase = await this.resolveLoadedTools(agentCode, agentVersion);
         const skillBundle = await this.resolveLoadedSkills(agentCode, userId, agentVersion);
         const loadedTools = assembleLoadedTools(loadedToolsBase);
@@ -681,11 +748,12 @@ export class AgentRuntime {
         const agentSystemPrompt = (await resolveAgentSystemPromptFromLoader(this.deps.graphLoader, agentCode, agentVersion)) ?? (typeof forkedCheckpoint.state.agentSystemPrompt === 'string'
             ? forkedCheckpoint.state.agentSystemPrompt
             : undefined);
+        const abortDeadlineMs = Date.now() + timeout;
         try {
             const forkState = { ...forkedCheckpoint.state };
             delete forkState.toolSpanId;
             delete forkState.hitlToolCallId;
-            const result = await this.deps.graphEngine.invoke(graph, {
+            const result = await runWithExecutionTrace({ abortSignal, abortDeadlineMs }, () => this.deps.graphEngine.invoke(graph, {
                 messages: Array.isArray(forkState.messages)
                     ? forkState.messages
                     : [],
@@ -701,7 +769,9 @@ export class AgentRuntime {
                 taskId: task.id,
                 traceId,
                 traceCollector: this.deps.traceCollector,
-            });
+                abortSignal,
+                abortDeadlineMs,
+            }));
             if (result.tokenUsage.perNode && result.tokenUsage.perNode.length > 0) {
                 await recordTaskTokenUsages(this.deps.taskService, task.id, result.tokenUsage.perNode);
             }
@@ -760,6 +830,23 @@ export class AgentRuntime {
             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);
+            const timedOut = err instanceof Error && /timed out/i.test(err.message);
+            if (timedOut) {
+                await this.deps.taskService.tryUpdateStatus(task.id, 'running', 'timeout');
+                return {
+                    newTraceId: traceId,
+                    originalTraceId,
+                    newTaskId: task.id,
+                    checkpointId,
+                    status: 'timeout',
+                    output: message,
+                    tokenUsage: {
+                        input_tokens,
+                        output_tokens,
+                        total_tokens: input_tokens + output_tokens,
+                    },
+                };
+            }
             await this.deps.taskService.updateStatus(task.id, 'failed', message);
             return {
                 newTraceId: traceId,
@@ -775,6 +862,9 @@ export class AgentRuntime {
                 },
             };
         }
+        finally {
+            this.deps.timeoutManager?.clearTimer(task.id);
+        }
     }
     async prepareHitlResume(taskId, response) {
         const task = await this.deps.taskService.getTask(taskId);

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

@@ -33,6 +33,7 @@ import { LongTermMemoryService } from '../memory/long-term-memory.service.js';
 export interface CreateDevRuntimeOptions {
     interceptors?: ExecutionInterceptors;
     compression?: {
+        /** 默认开。false 或 EAP_RUNTIME_COMPRESSION=0 关闭 */
         enabled?: boolean;
         persistCompressed?: boolean;
     };

package/dist/runtime/dev-runtime.js

@@ -31,7 +31,7 @@ import { createToolExecutor } from '../tool/create-tool-executor.js';
 import { DevToolExecutor } from '../tool/dev-tool-executor.js';
 import { captureOverlayJailOptions, withHarnessPrimitives } from '../tool/harness-primitives.js';
 import { AllowAllGovernanceClient, HttpGovernanceClient, } from '../governance/index.js';
-import { InvokeOrchestrator, StreamOrchestrator, SessionService, MessageService, SessionContextService, TaskService, LangGraphAdapter, HitlHandler, TokenBudgetService, TokenStatsService, TokenBudgetGuard, SessionBudgetChecker, createAgentRegistryClient, GraphLoader, } from '../index.js';
+import { InvokeOrchestrator, StreamOrchestrator, SessionService, MessageService, SessionContextService, TaskService, TaskTimeoutManager, LangGraphAdapter, HitlHandler, TokenBudgetService, TokenStatsService, TokenBudgetGuard, SessionBudgetChecker, createAgentRegistryClient, GraphLoader, } from '../index.js';
 import { createAgentRuntime } from './agent-runtime.js';
 import { assembleRuntimeStorage } from './storage-profile.js';
 import { createDevRagNodeDeps } from './dev-rag-deps.js';
@@ -139,6 +139,7 @@ export function createDevRuntime(options) {
     const sessionContextService = new SessionContextService(sessionService, messageService);
     // ── Task / Token ──
     const taskService = new TaskService();
+    const timeoutManager = new TaskTimeoutManager();
     const budgetService = new TokenBudgetService(storage.tokenBudgetRepository);
     const tokenStatsService = new TokenStatsService();
     const tokenBudgetGuard = new TokenBudgetGuard(budgetService);
@@ -202,6 +203,7 @@ export function createDevRuntime(options) {
         checkpointSaver: storage.checkpointSaver,
         sessionCompression,
         sessionMemory,
+        timeoutManager,
     });
     const streamOrchestrator = new StreamOrchestrator({
         sessionContextService,
@@ -215,6 +217,7 @@ export function createDevRuntime(options) {
         checkpointSaver: storage.checkpointSaver,
         sessionCompression,
         sessionMemory,
+        timeoutManager,
     });
     const agentRuntime = createAgentRuntime({
         invokeOrchestrator,
@@ -229,6 +232,7 @@ export function createDevRuntime(options) {
         turnLog,
         sessionCompression,
         sessionMemory,
+        timeoutManager,
     });
     return {
         invokeOrchestrator,
@@ -243,7 +247,10 @@ export function createDevRuntime(options) {
         taskService,
         tokenStatsService,
         graphLoader,
-        dispose: stopSkillWatch,
+        dispose: () => {
+            stopSkillWatch();
+            timeoutManager.dispose();
+        },
         diagnostics: {
             packageVersion: RUNTIME_PACKAGE_VERSION,
             deploymentContext,

package/dist/runtime/types.d.ts

@@ -78,7 +78,7 @@ export interface ForkResult {
     originalTraceId: string;
     newTaskId: string;
     checkpointId: string;
-    status: 'completed' | 'failed';
+    status: 'completed' | 'failed' | 'timeout';
     output: unknown;
     tokenUsage: {
         input_tokens: number;

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

@@ -1,11 +1,11 @@
 /**
- * execute/stream 装载历史后的压缩入口。默认关。
+ * execute/stream 装载历史后的压缩入口。默认开。
  *
- * C-1a 开关：options.enabled 或 EAP_RUNTIME_COMPRESSION=1
+ * C-1a 开关：默认开。options.enabled=false 或 EAP_RUNTIME_COMPRESSION=0 / false 关闭
  * C-1b 有窗口：用上一轮 llm_turns 的 used/window，不用字符估窗
  * C-1c 无窗口：只按轮次上限，不按窗压
- * C-1d 摘要只走 ILLMProvider；失败则不压缩
- * C-1e 默认只改本轮内存历史；persistCompressed 时必须能写 compression record
+ * C-1d 摘要只走 ILLMProvider；Echo 或未注入 provider 则不压；失败则不压缩
+ * C-1e 压成功且来自 Checkpoint 时，入图 `__set` 覆盖最新 Checkpoint messages；不回写 Message 库。persistCompressed 只写审计 record
  * C-1f 流式路径在摘要前产出 session.compression.started，结束后产出 completed。
  *      不新开「占用 ≥ 80% 再 compact」链；触发仍是 C-1b/c。
  */
@@ -15,6 +15,8 @@ import type { ILlmTurnLog } from '../debug/llm-turn-log.js';
 import { type CompressionRecord } from '../memory/message-compression.service.js';
 import type { ICompressionRepository } from '../memory/compression/compression.repository.js';
 import type { SessionCompressionCompletedPayload, SessionCompressionReason, SessionCompressionStartedPayload } from '../sse/events/session-events.js';
+/** 摘要 LLM 写入 llm_turns 的 nodeId。lastTurnUsage 跳过，避免用摘要占用触发下一轮压缩。 */
+export declare const COMPRESSION_TURN_NODE_ID = "session-compression";
 export interface SessionCompressionOptions {
     enabled?: boolean;
     persistCompressed?: boolean;

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

@@ -1,11 +1,11 @@
 /**
- * execute/stream 装载历史后的压缩入口。默认关。
+ * execute/stream 装载历史后的压缩入口。默认开。
  *
- * C-1a 开关：options.enabled 或 EAP_RUNTIME_COMPRESSION=1
+ * C-1a 开关：默认开。options.enabled=false 或 EAP_RUNTIME_COMPRESSION=0 / false 关闭
  * C-1b 有窗口：用上一轮 llm_turns 的 used/window，不用字符估窗
  * C-1c 无窗口：只按轮次上限，不按窗压
- * C-1d 摘要只走 ILLMProvider；失败则不压缩
- * C-1e 默认只改本轮内存历史；persistCompressed 时必须能写 compression record
+ * C-1d 摘要只走 ILLMProvider；Echo 或未注入 provider 则不压；失败则不压缩
+ * C-1e 压成功且来自 Checkpoint 时，入图 `__set` 覆盖最新 Checkpoint messages；不回写 Message 库。persistCompressed 只写审计 record
  * C-1f 流式路径在摘要前产出 session.compression.started，结束后产出 completed。
  *      不新开「占用 ≥ 80% 再 compact」链；触发仍是 C-1b/c。
  */
@@ -13,13 +13,17 @@ import { EchoLLMProvider } from '../providers/echo-llm-provider.js';
 import { COMPRESSION_WINDOW_FACTOR, countRounds, DEFAULT_COMPRESSION_CONFIG, } from '../memory/message-compression.service.js';
 import { SlidingWindowCompression } from '../memory/compression/sliding-window.compression.js';
 import { emitRuntimeLog } from '../common/logger.js';
+/** 摘要 LLM 写入 llm_turns 的 nodeId。lastTurnUsage 跳过，避免用摘要占用触发下一轮压缩。 */
+export const COMPRESSION_TURN_NODE_ID = 'session-compression';
 export function isRuntimeCompressionEnabled(explicit, env = process.env) {
     if (explicit === true)
         return true;
     if (explicit === false)
         return false;
     const raw = env.EAP_RUNTIME_COMPRESSION ?? env.EAP_COMPRESSION_ENABLED;
-    return raw === '1' || raw === 'true';
+    if (raw === '0' || raw === 'false')
+        return false;
+    return true;
 }
 export function isRuntimeCompressionPersist(explicit, env = process.env) {
     if (explicit === true)
@@ -164,7 +168,7 @@ async function runDecidedSessionCompression(params, decision) {
             maxRounds,
             preserveRecentRounds: options.preserveRecentRounds ?? DEFAULT_COMPRESSION_CONFIG.preserveRecentRounds,
             ...(decision.window ? { maxContextTokens: decision.window } : {}),
-        }, async (msgs) => summarizeWithProvider(provider, msgs, options.summaryModel));
+        }, async (msgs) => summarizeWithProvider(provider, msgs, options.summaryModel, params.sessionId));
         const compressed = await compression.compress(history);
         if (compressed === history) {
             return {
@@ -250,12 +254,17 @@ async function lastTurnUsage(turnLog, sessionId) {
     if (!turnLog || !sessionId)
         return undefined;
     const index = await turnLog.list(sessionId);
-    const last = index[index.length - 1];
-    if (!last || typeof last.used !== 'number')
-        return undefined;
-    return { used: last.used, window: last.window };
+    for (let i = index.length - 1; i >= 0; i--) {
+        const entry = index[i];
+        if (!entry || entry.nodeId === COMPRESSION_TURN_NODE_ID)
+            continue;
+        if (typeof entry.used !== 'number')
+            continue;
+        return { used: entry.used, window: entry.window };
+    }
+    return undefined;
 }
-async function summarizeWithProvider(provider, messages, summaryModel) {
+async function summarizeWithProvider(provider, messages, summaryModel, sessionId) {
     const body = messages
         .map((m) => `${m.role}: ${m.content}`)
         .join('\n')
@@ -272,6 +281,10 @@ async function summarizeWithProvider(provider, messages, summaryModel) {
         model,
         temperature: 0.2,
         maxTokens: 1024,
+        turn: {
+            sessionId,
+            nodeId: COMPRESSION_TURN_NODE_ID,
+        },
     });
     const text = response.content?.trim();
     if (!text) {

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, withCompressionSummaryModel, runSessionCompression, decideSessionCompression, iterSessionCompression, formatCompressionStartedMessage, formatCompressionCompletedMessage, isRuntimeCompressionEnabled, isRuntimeCompressionPersist, type SessionCompressionOptions, type RunSessionCompressionParams, type RunSessionCompressionResult, type SessionCompressionDecision, type SessionCompressionStep, } from './apply-session-compression.js';
+export { applySessionCompression, COMPRESSION_TURN_NODE_ID, 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, withCompressionSummaryModel, runSessionCompression, decideSessionCompression, iterSessionCompression, formatCompressionStartedMessage, formatCompressionCompletedMessage, isRuntimeCompressionEnabled, isRuntimeCompressionPersist, } from './apply-session-compression.js';
+export { applySessionCompression, COMPRESSION_TURN_NODE_ID, 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

@@ -99,6 +99,15 @@ export declare function appendTerminalToolResult(eventType: string, payload: unk
 }>): void;
 /** 历史 MessageRecord → GraphMessage（保留 tool 元数据与 reasoning） */
 export declare function messageRecordsToGraphMessages(records: MessageRecord[]): GraphMessage[];
+type CheckpointLoad = {
+    state?: {
+        messages?: unknown;
+    };
+    engineSnapshot?: Record<string, unknown>;
+} | null;
+type CheckpointLoader = {
+    load(threadId: string): Promise<CheckpointLoad>;
+};
 /**
  * 组装本轮送入 Graph 的 messages。
  *
@@ -107,21 +116,52 @@ export declare function messageRecordsToGraphMessages(records: MessageRecord[]):
  * - Session 级 Checkpoint 持有图内 messages（含 reasoning）
  * - 当 Checkpoint 已有历史时，只追加本轮 user，避免双重 append
  * - Checkpoint 缺失/无 messages 时，回退 Session 全量历史
+ * - 压缩针对 Checkpoint（无则 Session）中的模型上下文；压成功且来自 Checkpoint 时由调用方设 `replaceCheckpointMessages`
  */
 export declare function buildTurnGraphMessages(params: {
     history: GraphMessage[];
     input: string;
-    checkpointSaver?: {
-        load(threadId: string): Promise<{
-            state?: {
-                messages?: unknown;
-            };
-            engineSnapshot?: Record<string, unknown>;
-        } | null>;
-    } | null;
+    checkpointSaver?: CheckpointLoader | null;
     /** 与 LangGraph configurable.thread_id 一致（通常为 sessionId） */
     threadId?: string;
 }): Promise<GraphMessage[]>;
+/**
+ * 模型上下文：Checkpoint 已有 messages 用它，否则用 Session 历史。
+ * 压缩必须压这份列表，不能只压 Session 再在有 Checkpoint 时丢掉。
+ */
+export declare function loadModelFacingHistory(params: {
+    sessionHistory: GraphMessage[];
+    checkpointSaver?: CheckpointLoader | null;
+    threadId?: string;
+}): Promise<{
+    messages: GraphMessage[];
+    fromCheckpoint: boolean;
+}>;
+/**
+ * 压缩后的入图列表。来自 Checkpoint 且本轮压成功：整表 `__set` 覆盖，写入最新 Checkpoint。
+ * 未压且来自 Checkpoint：只追加本轮 user。无 Checkpoint：历史 + 本轮 user。
+ */
+export declare function turnGraphMessagesAfterCompression(params: {
+    history: GraphMessage[];
+    input: string;
+    fromCheckpoint: boolean;
+    compressed: boolean;
+}): {
+    messages: GraphMessage[];
+    replaceCheckpointMessages: boolean;
+};
+/**
+ * invoke 从 Graph 最终 messages 切本轮增量的起点。
+ * 覆盖（`__set`）或无 Checkpoint：入图数组长度（已含本轮 user）。
+ * 追加到已有 Checkpoint：模型上下文条数 + 入图条数（通常 1 条 user）。
+ */
+export declare function persistGraphMessageBaseline(params: {
+    fromCheckpoint: boolean;
+    replaceCheckpointMessages: boolean;
+    graphInputCount: number;
+    modelFacingCount: number;
+}): number;
+export declare function graphMessagesFromCheckpoint(cp: CheckpointLoad): GraphMessage[];
 /**
  * HITL 等待态写入 Task.output 的结构化载荷。
  * 冷恢复时优先读 `hitlThreadId`，避免误用 correlationId。
@@ -146,7 +186,7 @@ export declare function extractHitlThreadId(output: unknown): string | undefined
  * 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）。
+ * 组装本轮 Graph 输入请用 {@link loadModelFacingHistory} 与 {@link turnGraphMessagesAfterCompression}。
  */
 export declare function resolveCheckpointThreadId(params: {
     sessionId?: string;
@@ -199,3 +239,4 @@ export declare function persistTurnTranscript(messageService: MessageService | u
     assistantTokenCount?: number;
     assistantDurationMs?: number | null;
 }): Promise<void>;
+export {};