nolo-cli 0.67.0 → 0.68.0-alpha.2

package/{chunk-RCNQCXPM.js → chunk-5JTJG4OG.js}

@@ -8,7 +8,7 @@ import {
   readSseFrames,
   streamSseDataValues,
   throwIfChatCompletionStreamFailed
-} from "./chunk-M3HN7RBV.js";
+} from "./chunk-ZCVO2HLZ.js";
 import {
   clipToolText,
   hasImageInRuntimeMessages,
@@ -39,7 +39,7 @@ import {
   requiresBareImageUrl,
   resolveToolGuidedSections,
   toBareImageUrlMessages
-} from "./chunk-AVJQ4TDL.js";
+} from "./chunk-K5YS7FAN.js";
 import {
   convertMessagesToResponsesInput,
   extractReasoningFromResponseOutput,
@@ -80,16 +80,16 @@ import {
   EMPTY_ASSISTANT_REPAIR_PROMPT,
   getModelContextWindow,
   hasAssistantVisibleOutput,
+  normalizeUsage,
   readCacheCreationInputTokens,
   readCacheReadInputTokens,
   resolveEmptyAssistantFallbackMessage,
   resolveEmptyAssistantOutcome,
   resolveTruncatedReasoningTailLog
-} from "./chunk-HQ3ZLBSI.js";
+} from "./chunk-2XENHR3P.js";
 import {
   COMPACTION_SUMMARY_SCHEMA_VERSION,
   COMPACTION_SUMMARY_SYSTEM_PROMPT,
-  TOOL_STUB_TEXT,
   buildCompactionMetricsFromPlan,
   buildCompactionUserContent,
   clipCompactText,
@@ -102,8 +102,10 @@ import {
   observeMeaningfulProviderResponse,
   planCompression,
   planContextUsage,
+  resolveCompressionTriggerRatio,
+  truncateToolOutputForContext,
   withProviderCallTimingFields
-} from "./chunk-TDTSQVDX.js";
+} from "./chunk-3E67YXGD.js";
 import {
   estimateTokenCount
 } from "./chunk-PSCNXBYA.js";
@@ -12404,35 +12406,10 @@ function projectHistoryWithSummary(args2) {
     const found = bridged.findIndex((m) => m.id === args2.summarizedBeforeId);
     if (found !== -1) startIndex = found + 1;
   }
-  let stubEndIndex = -1;
-  if (args2.stubbedBeforeId) {
-    const found = bridged.findIndex((m) => m.id === args2.stubbedBeforeId);
-    if (found !== -1) stubEndIndex = found;
-  }
-  const projected = args2.history.slice(startIndex).map((message, i) => {
-    const absIndex = startIndex + i;
-    if (stubEndIndex !== -1 && message.role === "tool" && absIndex <= stubEndIndex) {
-      const content = typeof message.content === "string" ? TOOL_STUB_TEXT : Array.isArray(message.content) ? [{ type: "text", text: TOOL_STUB_TEXT }] : TOOL_STUB_TEXT;
-      return { ...message, content };
-    }
-    return message;
-  });
+  const projected = args2.history.slice(startIndex);
   const summaryMsg = args2.summary.trim().length > 0 ? [buildLocalSummaryHistoryMessage(args2.summary)] : [];
   return [...summaryMsg, ...projected];
 }
-function applyStoredStubToHistory(history, stubbedBeforeId) {
-  if (!stubbedBeforeId) return history;
-  const bridged = toPlanCompressionMessages(history);
-  const stubEndIndex = bridged.findIndex((m) => m.id === stubbedBeforeId);
-  if (stubEndIndex === -1) return history;
-  return history.map((message, i) => {
-    if (i <= stubEndIndex && message.role === "tool") {
-      const content = typeof message.content === "string" ? TOOL_STUB_TEXT : Array.isArray(message.content) ? [{ type: "text", text: TOOL_STUB_TEXT }] : TOOL_STUB_TEXT;
-      return { ...message, content };
-    }
-    return message;
-  });
-}
 var COLD_RESUME_IDLE_MS = 60 * 60 * 1e3;
 function isColdResume(history, nowMs) {
   for (let i = history.length - 1; i >= 0; i -= 1) {
@@ -12482,31 +12459,24 @@ async function maybeAutoCompactLocalHistory(args2) {
   }
   const existingSummary = typeof stored?.summary === "string" ? stored.summary : "";
   const summarizedBeforeId = typeof stored?.summarizedBeforeId === "string" ? stored.summarizedBeforeId : void 0;
-  const storedStubbedBeforeId = typeof stored?.stubbedBeforeId === "string" && stored.stubbedBeforeId ? stored.stubbedBeforeId : void 0;
   const contextWindow = typeof args2.contextWindow === "number" && Number.isFinite(args2.contextWindow) && args2.contextWindow > 0 ? args2.contextWindow : getModelContextWindow(args2.model ?? "");
-  const allMsgs = toPlanCompressionMessages(
-    applyStoredStubToHistory(history, storedStubbedBeforeId)
-  );
+  const allMsgs = toPlanCompressionMessages(history);
   const coldResume = isColdResume(history, args2.now?.() ?? Date.now());
   const plan = planCompression({
     allMsgs,
     summarizedBeforeId,
     summary: existingSummary,
     contextWindow,
-    // 防死亡螺旋：用 summary 长度作为上次压缩后的基线。如果新内容没让
-    // totalUsed 比 summary 本身增长超过 minNewTokens，不重复触发。
-    lastCompactedTokenCount: existingSummary ? estimateTokenCount(existingSummary) : void 0,
-    ...coldResume ? { force: true, reason: "context_budget" } : {},
+    ...coldResume ? { coldResume: true, reason: "cold_resume" } : {},
     ...args2.realContextUsagePercent !== void 0 ? { realContextUsagePercent: args2.realContextUsagePercent } : {}
   });
   const projectExisting = () => {
-    if (!existingSummary.trim() && !storedStubbedBeforeId) return unchanged();
+    if (!existingSummary.trim()) return unchanged();
     return {
       history: projectHistoryWithSummary({
         history,
         summary: existingSummary,
-        summarizedBeforeId,
-        stubbedBeforeId: storedStubbedBeforeId
+        summarizedBeforeId
       }),
       compressed: true,
       summaryGenerated: false
@@ -12515,49 +12485,6 @@ async function maybeAutoCompactLocalHistory(args2) {
   if (!plan.shouldCompress) {
     return projectExisting();
   }
-  if (plan.stub) {
-    const nextStubbedBeforeId = plan.stub.beforeId;
-    const changed = nextStubbedBeforeId !== storedStubbedBeforeId;
-    if (changed) {
-      try {
-        await adapter.saveDialogSummary({
-          dialogId,
-          summary: existingSummary,
-          summarizedBeforeId,
-          stubbedBeforeId: nextStubbedBeforeId,
-          // stub 档不改锚点切片 → 透传源校验字段保持读写对称
-          sourceHash: stored?.sourceHash,
-          sourceCount: stored?.sourceCount,
-          schemaVersion: COMPACTION_SUMMARY_SCHEMA_VERSION
-        });
-      } catch (error) {
-        console.warn("[localLoop] saveDialogSummary (stub) failed:", error);
-      }
-    }
-    const metrics = buildCompactionMetricsFromPlan({
-      reason: "tool_stub",
-      previousSummary: existingSummary,
-      plan,
-      newSummary: existingSummary
-    });
-    console.log(formatCompactionMetricsLog(metrics));
-    return {
-      history: projectHistoryWithSummary({
-        history,
-        summary: existingSummary,
-        summarizedBeforeId,
-        stubbedBeforeId: nextStubbedBeforeId
-      }),
-      compressed: true,
-      summaryGenerated: false,
-      metrics,
-      reason: "tool_stub",
-      savedTokens: plan.stub.savedTokens,
-      stubbedCount: plan.stub.stubbedCount,
-      beforeTokens: metrics.previousSummaryTokens + metrics.compressedTokens,
-      afterTokens: metrics.newSummaryTokens + metrics.retainedTokens
-    };
-  }
   try {
     const provider = await args2.resolveProvider();
     const msgsToCompress = plan.msgsToCompress;
@@ -12580,14 +12507,15 @@ async function maybeAutoCompactLocalHistory(args2) {
       console.warn(
         "[localLoop] auto-compaction produced empty summary; keeping prior projection"
       );
-      return projectExisting();
+      return {
+        ...projectExisting(),
+        failureMessage: "summary model returned empty content"
+      };
     }
     await adapter.saveDialogSummary({
       dialogId,
       summary: newSummary,
       summarizedBeforeId: plan.newSummarizedBeforeId,
-      // 新摘要已覆盖 stub 区间 → 清空 stubbedBeforeId
-      stubbedBeforeId: void 0,
       // 内容寻址失效检测：存锚点切片哈希与长度，供下轮载入校验。
       sourceHash: hashSummarySourceSlice(history, plan.newSummarizedBeforeId),
       sourceCount: (() => {
@@ -12622,7 +12550,8 @@ async function maybeAutoCompactLocalHistory(args2) {
     };
   } catch (error) {
     console.warn("[localLoop] auto-compaction failed:", error);
-    return projectExisting();
+    const failureMessage = error instanceof Error && error.message ? error.message.slice(0, 200) : String(error).slice(0, 200);
+    return { ...projectExisting(), failureMessage };
   }
 }
 
@@ -13572,6 +13501,7 @@ async function runLocalAgentTurn(input) {
       throw error;
     }
     loopTimingMark("loadDialogHistory", 0);
+    const canonicalHistory = history;
     const identityBlock = buildIdentityBlock({
       agentName: agentConfig.name,
       agentId: agentConfig.key,
@@ -13609,6 +13539,8 @@ async function runLocalAgentTurn(input) {
       }
       return resolvedProvider;
     };
+    const contextWindow = getModelContextWindow(agentConfig.model ?? "");
+    const compressionTriggerRatio = resolveCompressionTriggerRatio(contextWindow);
     let realContextUsagePercent;
     if (input.continueDialogId && typeof input.adapter.loadLastContextUsage === "function") {
       try {
@@ -13616,7 +13548,6 @@ async function runLocalAgentTurn(input) {
           input.continueDialogId
         );
         const inputTokens = usage?.inputTokens;
-        const contextWindow = getModelContextWindow(agentConfig.model);
         if (typeof inputTokens === "number" && Number.isFinite(inputTokens) && inputTokens > 0 && typeof contextWindow === "number" && Number.isFinite(contextWindow) && contextWindow > 0) {
           realContextUsagePercent = Math.min(
             1,
@@ -13628,6 +13559,21 @@ async function runLocalAgentTurn(input) {
       }
     }
     let compactionUsage;
+    const emitCompactionObservation = (compacted) => {
+      if (compacted.compressed || compacted.summaryGenerated || compacted.failureMessage) {
+        emitLoopEvent(observationBoundary, {
+          kind: "compaction",
+          atMs: Date.now(),
+          reason: compacted.reason ?? "context_budget",
+          summaryGenerated: compacted.summaryGenerated,
+          compressed: compacted.compressed,
+          ...compacted.failureMessage ? { failed: true, detail: compacted.failureMessage } : {},
+          ...compacted.beforeTokens !== void 0 ? { beforeTokens: compacted.beforeTokens } : {},
+          ...compacted.afterTokens !== void 0 ? { afterTokens: compacted.afterTokens } : {},
+          ...compacted.savedTokens !== void 0 ? { savedTokens: compacted.savedTokens } : {}
+        });
+      }
+    };
     try {
       const compacted = await maybeAutoCompactLocalHistory({
         adapter: input.adapter,
@@ -13639,19 +13585,7 @@ async function runLocalAgentTurn(input) {
       });
       history = compacted.history;
       compactionUsage = compacted.usage;
-      if (compacted.compressed || compacted.summaryGenerated) {
-        emitLoopEvent(observationBoundary, {
-          kind: "compaction",
-          atMs: Date.now(),
-          reason: compacted.reason ?? "context_budget",
-          summaryGenerated: compacted.summaryGenerated,
-          compressed: compacted.compressed,
-          ...compacted.beforeTokens !== void 0 ? { beforeTokens: compacted.beforeTokens } : {},
-          ...compacted.afterTokens !== void 0 ? { afterTokens: compacted.afterTokens } : {},
-          ...compacted.savedTokens !== void 0 ? { savedTokens: compacted.savedTokens } : {},
-          ...compacted.stubbedCount !== void 0 ? { stubbedCount: compacted.stubbedCount } : {}
-        });
-      }
+      emitCompactionObservation(compacted);
     } catch (error) {
       console.warn("[localLoop] auto-compaction unexpected error:", error);
     }
@@ -13663,7 +13597,6 @@ async function runLocalAgentTurn(input) {
     loopTimingMark("trimHistoryToContextBudget", 0);
     const hasContextBlocks = callerScopes.some((block) => block.content.trim()) || mergedContextBlockScopes.some((block) => block.content.trim());
     const promptMessageCount = agentConfig.prompt?.trim() || hasContextBlocks ? 1 : 0;
-    const turnStartIndex = promptMessageCount + history.length;
     const builtMessages = buildMessages({
       prompt: agentConfig.prompt,
       contextBlockScopes: mergedContextBlockScopes,
@@ -13694,6 +13627,17 @@ async function runLocalAgentTurn(input) {
         atMs: Date.now()
       });
     }
+    const rawTurnMessages = messages.slice(
+      promptMessageCount + history.length
+    );
+    const pushMessage = (msg) => {
+      messages.push(msg);
+      rawTurnMessages.push(msg);
+    };
+    const capToolMessage = (msg) => msg.role === "tool" && typeof msg.content === "string" ? {
+      ...msg,
+      content: truncateToolOutputForContext(msg.content, contextWindow)
+    } : msg;
     const userInputText = extractUserInputText(input.input);
     let toolCallCount = 0;
     let result4;
@@ -13708,6 +13652,50 @@ async function runLocalAgentTurn(input) {
     let skipFinalAppend = false;
     let partialContent = "";
     const progressGuard = createLocalLoopProgressGuard(input.progressGuardConfig);
+    const maybeCompactInLoop = async () => {
+      if (!input.continueDialogId) return;
+      if (!contextUsage) return;
+      const inputTokens = normalizeUsage(contextUsage).input_tokens;
+      if (!inputTokens || inputTokens <= 0) return;
+      const ratio = Math.min(1, inputTokens / contextWindow);
+      if (ratio < compressionTriggerRatio) return;
+      try {
+        const compacted = await maybeAutoCompactLocalHistory({
+          adapter: input.adapter,
+          dialogId: input.continueDialogId,
+          // canonical 坐标：store 全量历史 + 本轮消息的「持久化形态」
+          // （applyPersistedTurnInput 会把首条 user 消息换成 paste 展开形态）。
+          // 锚点与 sourceHash 都必须按持久化形态计算——否则下轮从 store
+          // 重载后重算 hash 必不匹配，摘要被判无效、白付一次摘要调用。
+          history: [
+            ...canonicalHistory,
+            ...applyPersistedTurnInput(
+              rawTurnMessages,
+              input.persistedInput,
+              input.persistedInputReference
+            )
+          ],
+          model: agentConfig.model,
+          resolveProvider: resolveProviderOnce,
+          realContextUsagePercent: ratio
+        });
+        if (compacted.usage) {
+          compactionUsage = addOutOfBandUsage(compactionUsage, compacted.usage);
+        }
+        if (compacted.compressed) {
+          messages = downgradeUnparsableToolCalls([
+            ...messages.slice(0, promptMessageCount),
+            ...prepareHistoryForNextTurn(
+              compacted.history,
+              input.adapter.host === "cli" ? input.contextReferenceResolver : void 0
+            )
+          ]).messages;
+        }
+        emitCompactionObservation(compacted);
+      } catch (error) {
+        console.warn("[localLoop] in-loop auto-compaction failed:", error);
+      }
+    };
     try {
       const provider = await resolveProviderOnce();
       const applyPendingInjections = () => {
@@ -13722,7 +13710,7 @@ async function runLocalAgentTurn(input) {
         let injected = false;
         for (const text of pending) {
           if (typeof text !== "string" || !text.trim()) continue;
-          messages.push({ role: "user", content: text });
+          pushMessage({ role: "user", content: text });
           injected = true;
         }
         return injected;
@@ -13904,13 +13892,14 @@ ${reasoningTailLog}
               content: result4.content,
               ...result4.reasoning_content ? { reasoning_content: result4.reasoning_content } : {}
             };
-            messages.push(assistantMessage);
+            pushMessage(assistantMessage);
             if (applyPendingInjections()) {
               loopTimingMark("roundEnd", round);
               round += 1;
               continue;
             }
             messages.pop();
+            rawTurnMessages.pop();
           }
           break;
         }
@@ -13930,7 +13919,9 @@ ${reasoningTailLog}
             }
           }
           if (hasInlineExecutedTools) {
-            messages.push(...blocksToOpenAiMessages(outputBlocks));
+            for (const blockMsg of blocksToOpenAiMessages(outputBlocks)) {
+              pushMessage(capToolMessage(blockMsg));
+            }
             skipFinalAppend = true;
             break;
           }
@@ -13938,7 +13929,7 @@ ${reasoningTailLog}
           continue;
         }
         toolCallCount += toolCalls.length;
-        messages.push({
+        pushMessage({
           role: "assistant",
           content: result4.content || null,
           ...result4.reasoning_content ? { reasoning_content: result4.reasoning_content } : {},
@@ -14114,17 +14105,19 @@ ${reasoningTailLog}
             metadata: toolResult.metadata
           });
           loopTimingMark("toolResultFormat", round);
-          messages.push({
-            role: "tool",
-            content: formatToolMessageContent({
+          pushMessage(
+            capToolMessage({
+              role: "tool",
+              content: formatToolMessageContent({
+                toolName,
+                content: toolResult.content,
+                metadata: observedMetadata
+              }),
+              tool_call_id: toolCall.id,
               toolName,
-              content: toolResult.content,
-              metadata: observedMetadata
-            }),
-            tool_call_id: toolCall.id,
-            toolName,
-            ...observedMetadata ? { tool_result_metadata: observedMetadata } : {}
-          });
+              ...observedMetadata ? { tool_result_metadata: observedMetadata } : {}
+            })
+          );
         }
         const toolExecutionGuardVerdict = progressGuard.observeToolExecution(
           toolCalls,
@@ -14146,6 +14139,7 @@ ${reasoningTailLog}
           };
           break;
         }
+        await maybeCompactInLoop();
         loopTimingMark("roundEnd", round);
         round += 1;
       }
@@ -14154,7 +14148,7 @@ ${reasoningTailLog}
     }
     if (loopError) {
       const turnMessages2 = applyPersistedTurnInput(
-        messages.slice(turnStartIndex),
+        rawTurnMessages,
         input.persistedInput,
         input.persistedInputReference
       );
@@ -14192,7 +14186,7 @@ ${reasoningTailLog}
         result4.emptyAssistantFallbackReason,
         result4.reasoning_content
       );
-      messages.push({
+      pushMessage({
         role: "assistant",
         content: result4.content,
         // 与中间轮(:561)一致带上 reasoning_content,让 saveTurn 持久化思维链,
@@ -14202,7 +14196,7 @@ ${reasoningTailLog}
       });
     }
     const turnMessages = applyPersistedTurnInput(
-      messages.slice(turnStartIndex),
+      rawTurnMessages,
       input.persistedInput,
       input.persistedInputReference
     );