jinzd-ai-cli 0.4.185 → 0.4.187

package/dist/index.js

@@ -15,7 +15,7 @@ import {
   saveDevState,
   sessionHasMeaningfulContent,
   setupProxy
-} from "./chunk-U7P5A3MJ.js";
+} from "./chunk-ZTBUTA24.js";
 import {
   ToolExecutor,
   ToolRegistry,
@@ -34,10 +34,10 @@ import {
   spawnAgentContext,
   theme,
   undoStack
-} from "./chunk-GOS4DWW5.js";
+} from "./chunk-Q6BSUIDV.js";
 import "./chunk-HDSKW7Q3.js";
 import "./chunk-ZWVIDFGY.js";
-import "./chunk-GSRXKHZ7.js";
+import "./chunk-4SZ6X47A.js";
 import {
   SessionManager,
   getContentText
@@ -54,47 +54,35 @@ import {
   getConfigDirUsage,
   listRecentCrashes,
   writeCrashLog
-} from "./chunk-ZO4LKUDM.js";
+} from "./chunk-43T4MY5B.js";
 import {
-  BudgetWarner,
   CONTENT_ONLY_STREAM_REMINDER,
-  ContextPressureMonitor,
-  EmptyResponseGuard,
-  FreeRoundTracker,
-  HALLUCINATION_CORRECTION_MESSAGE,
   ProviderRegistry,
   TEE_FINAL_USER_NUDGE,
   TOOL_CALL_REMINDER,
   ThinkTagFilter,
   accumulateUsage,
-  buildPhantomCorrectionMessage,
   buildRoundBudgetHint,
-  buildRoundsExhaustedPrompt,
-  buildUserStopMessage,
   buildWriteRoundReminder,
   consumeToolCallStream,
   detectMetaNarration,
   detectPseudoToolCalls,
-  detectsHallucinatedFileOp,
-  extractBashCommands,
   extractWrittenFilePaths,
-  findPhantomClaims,
-  hadPreviousWriteToolCalls,
   looksLikeDocumentBody,
+  runAgentLoop,
   stripPseudoToolCalls,
-  stripToolCallReminder,
-  summarizeRecentTools
-} from "./chunk-5LK7H45B.js";
+  stripToolCallReminder
+} from "./chunk-IQ7JE43O.js";
 import {
   getStatsSnapshot,
   getTopFailingTools,
   getTopUsedTools,
   installFlushOnExit
-} from "./chunk-GH32XE5K.js";
+} from "./chunk-MVK25WZW.js";
 import "./chunk-HIU2SH4V.js";
 import {
   ConfigManager
-} from "./chunk-FHZ2LKM5.js";
+} from "./chunk-UAJKGLRV.js";
 import {
   AuthError,
   ProviderError,
@@ -121,7 +109,7 @@ import {
   SKILLS_DIR_NAME,
   VERSION,
   buildUserIdentityPrompt
-} from "./chunk-PMZCQAJL.js";
+} from "./chunk-SOWBY545.js";
 import {
   formatGitContextForPrompt,
   getGitContext,
@@ -1836,7 +1824,7 @@ No tools match "${filter}".
           const { join: join5 } = await import("path");
           const { existsSync: existsSync5 } = await import("fs");
           const { getGitRoot: getGitRoot2 } = await import("./git-context-7KIP4X2V.js");
-          const { MCP_PROJECT_CONFIG_NAME: MCP_PROJECT_CONFIG_NAME2 } = await import("./constants-RJDN7GOH.js");
+          const { MCP_PROJECT_CONFIG_NAME: MCP_PROJECT_CONFIG_NAME2 } = await import("./constants-XEL5347E.js");
           const { approveProject, hashMcpFile } = await import("./project-trust-IFM7FXEV.js");
           const cwd = process.cwd();
           const projectRoot = getGitRoot2(cwd) ?? cwd;
@@ -2897,7 +2885,7 @@ ${hint}` : "")
       usage: "/test [command|filter]",
       async execute(args, ctx) {
         try {
-          const { executeTests } = await import("./run-tests-37RHYYD4.js");
+          const { executeTests } = await import("./run-tests-Z7IGVS2W.js");
           const argStr = args.join(" ").trim();
           let testArgs = {};
           if (argStr) {
@@ -6391,7 +6379,6 @@ Session '${this.resumeSessionId}' not found.
       toolDefs = toolDefs.filter((t) => !this.blockedTools.has(t.name));
     }
     const apiMessages = [...messages];
-    const extraMessages = [];
     const maxToolRounds = this.maxToolRoundsOverride ?? this.config.get("maxToolRounds") ?? DEFAULT_MAX_TOOL_ROUNDS;
     const autoPauseIntervalRaw = this.config.get("autoPauseInterval");
     const autoPauseInterval = typeof autoPauseIntervalRaw === "number" ? autoPauseIntervalRaw : DEFAULT_AUTO_PAUSE_INTERVAL;
@@ -6404,55 +6391,175 @@ ${mcpBudgetNote}` : "");
     const modelParams = this.getModelParams(effectiveModel);
     const useStreaming = this.config.get("ui").streaming;
     const spinner = this.renderer.showSpinner("Thinking...");
-    const roundUsage = { inputTokens: 0, outputTokens: 0, cacheCreationTokens: 0, cacheReadTokens: 0 };
+    const usage = { inputTokens: 0, outputTokens: 0, cacheCreationTokens: 0, cacheReadTokens: 0 };
     const supportsStreamingTools = useStreaming && typeof provider.chatWithToolsStream === "function";
     let lastToolCallSignature = "";
     let repeatedToolCallCount = 0;
-    const roundToolHistory = [];
-    const budgetWarner = new BudgetWarner(maxToolRounds);
-    const emptyGuard = new EmptyResponseGuard();
-    const ctxMonitor = new ContextPressureMonitor();
-    const freeRounds = new FreeRoundTracker();
     this.setupInterjectionListener();
     try {
-      for (let round = 0; round < maxToolRounds; round++) {
-        this.toolExecutor.setRoundInfo(round + 1, maxToolRounds);
-        if (this.toolExecutor.pendingSlashCommand) {
-          const cmd = this.toolExecutor.pendingSlashCommand;
-          this.toolExecutor.pendingSlashCommand = null;
-          if (cmd === "/exit" || cmd === "/quit" || cmd === "/q") {
-            spinner.stop();
-            process.stdout.write(theme.warning(`\u26A1 ${cmd} \u2014 stopping agentic loop
-`));
-            this.teardownInterjectionListener();
-            return;
+      const loopResult = await runAgentLoop({
+        maxToolRounds,
+        autoPauseInterval,
+        planMode: this.planMode,
+        providerId: this.currentProvider,
+        toolDefs,
+        usage,
+        // ── 模型调用：流式/非流式 + retry/fallback 包装（disabled-by-default，v0.4.144+）──
+        callModel: async (_round, extraMessages) => {
+          const chatRequest = {
+            messages: apiMessages,
+            model: effectiveModel,
+            systemPrompt,
+            systemPromptVolatile,
+            stream: false,
+            temperature: modelParams.temperature,
+            maxTokens: modelParams.maxTokens,
+            timeout: modelParams.timeout,
+            thinking: modelParams.thinking,
+            thinkingBudget: modelParams.thinkingBudget,
+            ...extraMessages.length > 0 ? { _extraMessages: extraMessages } : {}
+          };
+          const fallbackOpts = this.getFallbackOptions(spinner);
+          if (supportsStreamingTools) {
+            const streamAc = this.setupStreamInterrupt();
+            try {
+              const streamGen = withFallbackStream(
+                this.currentProvider,
+                effectiveModel,
+                this.providers,
+                fallbackOpts,
+                (p, m) => {
+                  const tc = p;
+                  if (typeof tc.chatWithToolsStream !== "function") {
+                    throw new Error(`provider ${p.info.id} does not support streaming tool calls`);
+                  }
+                  return tc.chatWithToolsStream({ ...chatRequest, model: m, signal: streamAc.signal }, toolDefs);
+                }
+              );
+              const streamResult = await this.consumeToolStream(streamGen, spinner);
+              if (streamResult.toolCalls.length > 0) {
+                return {
+                  toolCalls: streamResult.toolCalls,
+                  usage: streamResult.usage,
+                  reasoningContent: streamResult.reasoningContent
+                };
+              }
+              return {
+                content: streamResult.textContent,
+                usage: streamResult.usage,
+                reasoningContent: streamResult.reasoningContent,
+                finishReason: streamResult.finishReason,
+                alreadyRendered: true
+              };
+            } finally {
+              this.teardownStreamInterrupt();
+            }
           }
-          process.stdout.write(theme.warning(`\u26A1 Command "${cmd}" \u2014 injected as message
-`));
-          extraMessages.push({ role: "user", content: cmd });
-        }
-        const budgetWarning = budgetWarner.check(maxToolRounds - round);
-        if (budgetWarning) {
-          extraMessages.push({ role: "user", content: budgetWarning.injectMessage });
-          if (budgetWarning.displayMessage) {
-            const paint = budgetWarning.level === "critical" ? theme.error : theme.warning;
-            process.stdout.write(paint(`  ${budgetWarning.displayMessage}
-`));
+          const result = await withFallback(
+            this.currentProvider,
+            effectiveModel,
+            this.providers,
+            fallbackOpts,
+            (p, m) => p.chatWithTools({ ...chatRequest, model: m }, toolDefs)
+          );
+          return result;
+        },
+        // ── 轮次耗尽总结：空工具列表强制纯文本 ──
+        callSummary: async (summaryExtra) => {
+          spinner.stop();
+          spinner.start("Generating summary...");
+          try {
+            const summaryResult = await provider.chatWithTools(
+              {
+                messages: apiMessages,
+                model: effectiveModel,
+                systemPrompt,
+                systemPromptVolatile,
+                stream: false,
+                temperature: modelParams.temperature,
+                maxTokens: modelParams.maxTokens,
+                timeout: modelParams.timeout,
+                thinking: modelParams.thinking,
+                thinkingBudget: modelParams.thinkingBudget,
+                _extraMessages: summaryExtra
+              },
+              []
+              // 不提供任何工具，强制 AI 返回纯文本
+            );
+            spinner.stop();
+            return "content" in summaryResult ? { content: summaryResult.content, usage: summaryResult.usage } : { usage: summaryResult.usage };
+          } catch (err) {
+            spinner.stop();
+            throw err;
           }
-        }
-        if (this._userInterjection) {
+        },
+        executeTools: async (toolCalls, extraMessages) => {
+          spinner.stop();
+          askUserContext.rl = this.rl;
+          googleSearchContext.configManager = this.config;
+          streamToFileContext.provider = provider;
+          streamToFileContext.model = effectiveModel;
+          streamToFileContext.systemPrompt = systemPromptVolatile ? `${systemPrompt}
+
+---
+
+${systemPromptVolatile}` : systemPrompt;
+          streamToFileContext.messages = apiMessages;
+          streamToFileContext.extraMessages = extraMessages;
+          streamToFileContext.temperature = modelParams.temperature;
+          streamToFileContext.timeout = modelParams.timeout;
+          spawnAgentContext.provider = provider;
+          spawnAgentContext.model = effectiveModel;
+          spawnAgentContext.systemPrompt = systemPromptVolatile ? `${systemPrompt}
+
+---
+
+${systemPromptVolatile}` : systemPrompt;
+          spawnAgentContext.modelParams = modelParams;
+          spawnAgentContext.configManager = this.config;
+          ToolExecutor.currentMessageIndex = session.messages.length;
+          return this.toolExecutor.executeAll(toolCalls);
+        },
+        buildToolResultMessages: (toolCalls, results, reasoningContent) => provider.buildToolResultMessages(toolCalls, results, reasoningContent),
+        getContextWindow: () => this.getContextWindowSize(),
+        estimateRequestTokens: (extraMessages) => this.estimateRequestTokens(systemPrompt, extraMessages),
+        isInterrupted: () => isInterrupted(),
+        pollInterjection: () => {
+          if (!this._userInterjection) return null;
           const msg = this._userInterjection;
           this._userInterjection = null;
           process.stdout.write(theme.warning(`\u26A1 Interjection: "${msg}"
 `));
-          extraMessages.push({ role: "user", content: msg });
-        }
-        const ctxWindow = this.getContextWindowSize();
-        if (ctxWindow > 0) {
-          const reqTokens = this.estimateRequestTokens(systemPrompt, extraMessages);
-          const pressure = ctxMonitor.check(reqTokens, ctxWindow);
+          return msg;
+        },
+        onRoundStart: (round) => {
+          this.toolExecutor.setRoundInfo(round + 1, maxToolRounds);
+        },
+        beforeRound: (_round, extraMessages) => {
+          if (this.toolExecutor.pendingSlashCommand) {
+            const cmd = this.toolExecutor.pendingSlashCommand;
+            this.toolExecutor.pendingSlashCommand = null;
+            if (cmd === "/exit" || cmd === "/quit" || cmd === "/q") {
+              spinner.stop();
+              process.stdout.write(theme.warning(`\u26A1 ${cmd} \u2014 stopping agentic loop
+`));
+              return "stop";
+            }
+            process.stdout.write(theme.warning(`\u26A1 Command "${cmd}" \u2014 injected as message
+`));
+            extraMessages.push({ role: "user", content: cmd });
+          }
+        },
+        onBudgetWarning: (warning) => {
+          if (warning.displayMessage) {
+            const paint = warning.level === "critical" ? theme.error : theme.warning;
+            process.stdout.write(paint(`  ${warning.displayMessage}
+`));
+          }
+        },
+        onContextPressure: (pressure, ctxWindow, round) => {
+          spinner.stop();
           if (pressure.action === "abort") {
-            spinner.stop();
             process.stderr.write(
               theme.error(
                 `
@@ -6468,16 +6575,7 @@ ${mcpBudgetNote}` : "");
 `
               )
             );
-            if (roundUsage.inputTokens > 0 || roundUsage.outputTokens > 0) {
-              this.addSessionUsage(roundUsage, effectiveModel);
-              session.addTokenUsage(roundUsage);
-              if (this.shouldShowTokens()) {
-                this.renderer.renderUsage(roundUsage, this.sessionTokenUsage);
-              }
-            }
-            return;
-          } else if (pressure.action === "warn") {
-            spinner.stop();
+          } else {
             process.stdout.write(
               theme.warning(
                 `
@@ -6485,492 +6583,142 @@ ${mcpBudgetNote}` : "");
 `
               )
             );
-            extraMessages.push({ role: "user", content: pressure.injectMessage });
             spinner.start(`Thinking... (round ${round + 1}/${maxToolRounds})`);
           }
-        }
-        let result;
-        let alreadyRendered = false;
-        const chatRequest = {
-          messages: apiMessages,
-          model: effectiveModel,
-          systemPrompt,
-          systemPromptVolatile,
-          stream: false,
-          temperature: modelParams.temperature,
-          maxTokens: modelParams.maxTokens,
-          timeout: modelParams.timeout,
-          thinking: modelParams.thinking,
-          thinkingBudget: modelParams.thinkingBudget,
-          ...extraMessages.length > 0 ? { _extraMessages: extraMessages } : {}
-        };
-        if (supportsStreamingTools) {
-          const streamAc = this.setupStreamInterrupt();
-          try {
-            const fallbackOpts = this.getFallbackOptions(spinner);
-            const streamGen = withFallbackStream(
-              this.currentProvider,
-              effectiveModel,
-              this.providers,
-              fallbackOpts,
-              (p, m) => {
-                const tc = p;
-                if (typeof tc.chatWithToolsStream !== "function") {
-                  throw new Error(`provider ${p.info.id} does not support streaming tool calls`);
-                }
-                return tc.chatWithToolsStream({ ...chatRequest, model: m, signal: streamAc.signal }, toolDefs);
-              }
-            );
-            const streamResult = await this.consumeToolStream(streamGen, spinner);
-            if (streamResult.toolCalls.length > 0) {
-              const toolCalls = streamResult.toolCalls;
-              if (streamResult.rawContent) {
-                toolCalls._rawContent = streamResult.rawContent;
-              }
-              if (streamResult.textContent) {
-                toolCalls._streamedText = streamResult.textContent;
-              }
-              result = {
-                toolCalls,
-                usage: streamResult.usage,
-                ...streamResult.reasoningContent ? { reasoningContent: streamResult.reasoningContent } : {}
-              };
-            } else {
-              result = {
-                content: streamResult.textContent,
-                usage: streamResult.usage,
-                ...streamResult.reasoningContent ? { reasoningContent: streamResult.reasoningContent } : {},
-                ...streamResult.finishReason ? { finishReason: streamResult.finishReason } : {}
-              };
-              alreadyRendered = true;
-            }
-          } finally {
-            this.teardownStreamInterrupt();
-          }
-        } else {
-          const fallbackOpts = this.getFallbackOptions(spinner);
-          result = await withFallback(
-            this.currentProvider,
-            effectiveModel,
-            this.providers,
-            fallbackOpts,
-            (p, m) => p.chatWithTools({ ...chatRequest, model: m }, toolDefs)
-          );
-        }
-        accumulateUsage(roundUsage, result.usage);
-        if ("content" in result) {
-          const hasWriteTools = toolDefs.some((t) => t.name === "write_file" || t.name === "edit_file");
-          const alreadyWrote = hadPreviousWriteToolCalls(extraMessages);
-          const coarseHallucination = !this.planMode && hasWriteTools && !alreadyWrote && !!result.content && detectsHallucinatedFileOp(result.content);
-          const phantomPaths = (coarseHallucination || alreadyWrote) && !this.planMode && hasWriteTools && result.content ? findPhantomClaims(result.content, extraMessages) : [];
-          const bashRanThisTurn = extractBashCommands(extraMessages).length > 0;
-          const coarseShouldFire = coarseHallucination && !bashRanThisTurn;
-          if ((phantomPaths.length > 0 || coarseShouldFire) && round < maxToolRounds - 1) {
-            const providerName = this.currentProvider;
-            const detail = phantomPaths.length > 0 ? ` phantom files: ${phantomPaths.join(", ")}` : "";
-            process.stderr.write(
-              `[${providerName}] \u26A0 Hallucinated completion detected (AI claimed file was written but no tool was called), forcing retry...${detail}
+        },
+        onHallucinationRetry: ({ phantomPaths, round, alreadyRendered }) => {
+          const detail = phantomPaths.length > 0 ? ` phantom files: ${phantomPaths.join(", ")}` : "";
+          process.stderr.write(
+            `[${this.currentProvider}] \u26A0 Hallucinated completion detected (AI claimed file was written but no tool was called), forcing retry...${detail}
 `
-            );
-            if (alreadyRendered) {
-              process.stdout.write("\n");
-            }
-            const correctionMsg = phantomPaths.length > 0 ? buildPhantomCorrectionMessage(phantomPaths) : HALLUCINATION_CORRECTION_MESSAGE;
-            const reasoningField = "reasoningContent" in result && result.reasoningContent ? { reasoning_content: result.reasoningContent } : this.currentProvider === "deepseek" ? { reasoning_content: "" } : {};
-            extraMessages.push(
-              { role: "assistant", content: result.content, ...reasoningField },
-              { role: "user", content: correctionMsg }
-            );
-            spinner.start(`Retrying... (round ${round + 2}/${maxToolRounds})`);
-            continue;
+          );
+          if (alreadyRendered) {
+            process.stdout.write("\n");
           }
-          if (!result.content || result.content.trim() === "") {
-            const fr = "finishReason" in result ? result.finishReason : void 0;
-            const decision = emptyGuard.onEmpty(round < maxToolRounds - 1, fr);
-            if (decision.action === "nudge") {
-              spinner.stop();
-              if (alreadyRendered) process.stdout.write("\n");
-              process.stderr.write(theme.warning(`${decision.displayMessage}
+          spinner.start(`Retrying... (round ${round + 2}/${maxToolRounds})`);
+        },
+        onEmptyResponse: (decision, { alreadyRendered, round }) => {
+          spinner.stop();
+          if (alreadyRendered) process.stdout.write("\n");
+          if (decision.action === "nudge") {
+            process.stderr.write(theme.warning(`${decision.displayMessage}
 `));
-              extraMessages.push({ role: "user", content: decision.injectMessage });
-              spinner.start(`Retrying... (round ${round + 2}/${maxToolRounds})`);
-              continue;
-            }
-            spinner.stop();
-            if (alreadyRendered) process.stdout.write("\n");
-            process.stderr.write(
-              theme.error(`
+            spinner.start(`Retrying... (round ${round + 2}/${maxToolRounds})`);
+          } else {
+            process.stderr.write(theme.error(`
 ${decision.displayMessage}
-`)
-            );
+`));
             process.stderr.write(
               theme.dim(`  ${decision.hint}
   Try: /compact, /clear, or /model to switch.
 
 `)
             );
-            if (roundUsage.inputTokens > 0 || roundUsage.outputTokens > 0) {
-              this.addSessionUsage(roundUsage, effectiveModel);
-              session.addTokenUsage(roundUsage);
-              if (this.shouldShowTokens()) {
-                this.renderer.renderUsage(roundUsage, this.sessionTokenUsage);
-              }
-            }
-            return;
           }
-          emptyGuard.onNonEmpty();
+        },
+        onFinalContent: async (content, { reasoningContent, alreadyRendered }) => {
           spinner.stop();
-          const finalContent = result.content;
           if (!alreadyRendered) {
             if (useStreaming) {
               const streamAc = this.setupStreamInterrupt();
               try {
-                await this.renderer.renderContentAsStream(finalContent, { signal: streamAc.signal });
+                await this.renderer.renderContentAsStream(content, { signal: streamAc.signal });
               } finally {
                 this.teardownStreamInterrupt();
               }
             } else {
-              this.renderer.renderResponse(finalContent);
-            }
-          } else {
-            if (finalContent.trim()) {
-              process.stdout.write("\n\n");
+              this.renderer.renderResponse(content);
             }
+          } else if (content.trim()) {
+            process.stdout.write("\n\n");
           }
-          lastResponseStore.content = finalContent;
-          const finalReasoning = "reasoningContent" in result ? result.reasoningContent : void 0;
+          lastResponseStore.content = content;
           session.addMessage({
             role: "assistant",
-            content: finalContent,
+            content,
             timestamp: /* @__PURE__ */ new Date(),
-            ...finalReasoning ? { reasoningContent: finalReasoning } : {}
+            ...reasoningContent ? { reasoningContent } : {}
           });
-          this.events.emit("message.after", { content: finalContent });
-          if (roundUsage.inputTokens > 0 || roundUsage.outputTokens > 0) {
-            this.addSessionUsage(roundUsage, effectiveModel);
-            session.addTokenUsage(roundUsage);
-            if (this.shouldShowTokens()) {
-              this.renderer.renderUsage(roundUsage, this.sessionTokenUsage);
-            }
-          }
-          return;
-        }
-        spinner.stop();
-        const saveLastResponseCall = result.toolCalls.find((tc) => tc.name === "save_last_response");
-        if (saveLastResponseCall) {
-          const saveToFile = String(saveLastResponseCall.arguments["path"] ?? "");
-          if (!saveToFile) {
-          } else {
-            const teeAc = this.setupStreamInterrupt();
-            try {
-              const teeSystemPrompt = stripToolCallReminder(systemPrompt ?? "") + CONTENT_ONLY_STREAM_REMINDER;
-              const teeExtraMessages = extraMessages.length > 0 ? [...extraMessages, { role: "user", content: TEE_FINAL_USER_NUDGE }] : [{ role: "user", content: TEE_FINAL_USER_NUDGE }];
-              const genStream = provider.chatStream({
-                messages: apiMessages,
-                model: effectiveModel,
-                systemPrompt: teeSystemPrompt,
-                systemPromptVolatile,
-                stream: true,
-                temperature: modelParams.temperature,
-                maxTokens: modelParams.maxTokens,
-                timeout: modelParams.timeout,
-                thinking: modelParams.thinking,
-                thinkingBudget: modelParams.thinkingBudget,
-                signal: teeAc.signal,
-                _extraMessages: teeExtraMessages
-              });
-              const teeShowTokens = this.shouldShowTokens();
-              let genContent;
-              let genUsage;
-              let teeTokShown = false;
-              try {
-                const teeResult = await this.renderer.renderStream(
-                  genStream,
-                  { saveToFile, showTokens: teeShowTokens, sessionTotal: teeShowTokens ? { ...this.sessionTokenUsage } : void 0, signal: teeAc.signal }
-                );
-                genContent = teeResult.content;
-                genUsage = teeResult.usage;
-                teeTokShown = teeResult.tokensShown;
-              } catch (teeErr) {
-                try {
-                  unlinkSync2(saveToFile);
-                } catch {
-                }
-                const errMsg = teeErr instanceof Error ? teeErr.message : String(teeErr);
-                process.stdout.write(theme.error(
-                  `
-  \u2717 tee stream failed: ${errMsg}
-  ${saveToFile} (partial) was deleted. Asking model to retry.
-
-`
-                ));
-                const errorResults = result.toolCalls.map((tc) => ({
-                  callId: tc.id,
-                  content: tc.name === "save_last_response" ? `[save_last_response failed] streaming was interrupted: ${errMsg}. ${saveToFile} was NOT saved. Retry \u2014 and consider producing a more compact output (split very large reports across multiple save_last_response calls if the previous attempt timed out).` : `[skipped: save_last_response failed]`,
-                  isError: tc.name === "save_last_response"
-                }));
-                const reasoningContent3 = "reasoningContent" in result ? result.reasoningContent : void 0;
-                const newMsgs3 = provider.buildToolResultMessages(result.toolCalls, errorResults, reasoningContent3);
-                extraMessages.push(...newMsgs3);
-                continue;
-              }
-              const metaMatch = detectMetaNarration(genContent);
-              if (metaMatch) {
-                try {
-                  unlinkSync2(saveToFile);
-                } catch {
-                }
-                process.stdout.write(theme.error(
-                  `
-  \u2717 Rejected save: response was meta-narration / leaked reasoning, not document body (matched: ${metaMatch})
-  ${saveToFile} was deleted; asking model to retry.
-
-`
-                ));
-                const errorResults = result.toolCalls.map((tc) => ({
-                  callId: tc.id,
-                  content: tc.name === "save_last_response" ? `[save_last_response REJECTED] Your output was internal reasoning / meta-narration about the task (e.g. "Let me re-read\u2026", "the user is asking me to\u2026") instead of the requested document body. ${saveToFile} was NOT saved.
-
-This fresh stream has NO tools. Produce ONLY the document body: start with a markdown heading like "# \u5BA1\u8BA1\u62A5\u544A" / "# Audit Report" and write the full content. Do NOT narrate that you will produce the document \u2014 produce it.` : `[skipped: save_last_response was rejected and other parallel calls are abandoned]`,
-                  isError: tc.name === "save_last_response"
-                }));
-                const reasoningContent3 = "reasoningContent" in result ? result.reasoningContent : void 0;
-                const newMsgs3 = provider.buildToolResultMessages(result.toolCalls, errorResults, reasoningContent3);
-                extraMessages.push(...newMsgs3);
-                if (genUsage) {
-                  roundUsage.inputTokens += genUsage.inputTokens;
-                  roundUsage.outputTokens += genUsage.outputTokens;
-                  roundUsage.cacheCreationTokens += genUsage.cacheCreationTokens ?? 0;
-                  roundUsage.cacheReadTokens += genUsage.cacheReadTokens ?? 0;
-                }
-                continue;
-              }
-              const pseudoMatch = detectPseudoToolCalls(genContent);
-              if (pseudoMatch) {
-                const cleaned = stripPseudoToolCalls(genContent);
-                if (looksLikeDocumentBody(cleaned)) {
-                  try {
-                    writeFileSync2(saveToFile, cleaned, "utf-8");
-                    process.stdout.write(theme.warning(
-                      `
-  \u26A0 Salvaged save: stripped pseudo-tool-call markup (matched: ${pseudoMatch})
-  ${saveToFile} now contains the cleaned document (${cleaned.length} chars; was ${genContent.length}).
-
-`
-                    ));
-                    lastResponseStore.content = cleaned;
-                    if (genUsage) {
-                      roundUsage.inputTokens += genUsage.inputTokens;
-                      roundUsage.outputTokens += genUsage.outputTokens;
-                      roundUsage.cacheCreationTokens += genUsage.cacheCreationTokens ?? 0;
-                      roundUsage.cacheReadTokens += genUsage.cacheReadTokens ?? 0;
-                    }
-                    session.addMessage({ role: "assistant", content: cleaned, timestamp: /* @__PURE__ */ new Date() });
-                    this.events.emit("message.after", { content: cleaned });
-                    const lines2 = cleaned.split("\n").length;
-                    const bytes2 = Buffer.byteLength(cleaned, "utf-8");
-                    const okResults = result.toolCalls.map((tc) => ({
-                      callId: tc.id,
-                      content: tc.name === "save_last_response" ? `File saved (with cleanup): ${saveToFile} (${lines2} lines, ${bytes2} bytes; pseudo-tool-call markup was stripped before save)` : `[skipped: file already saved by tee streaming]`,
-                      isError: false
-                    }));
-                    const reasoningContent4 = "reasoningContent" in result ? result.reasoningContent : void 0;
-                    const newMsgs4 = provider.buildToolResultMessages(result.toolCalls, okResults, reasoningContent4);
-                    extraMessages.push(...newMsgs4);
-                    if (roundUsage.inputTokens > 0 || roundUsage.outputTokens > 0) {
-                      this.addSessionUsage(roundUsage, effectiveModel);
-                      session.addTokenUsage(roundUsage);
-                      if (teeShowTokens && !teeTokShown) {
-                        this.renderer.renderUsage(roundUsage, this.sessionTokenUsage);
-                      }
-                    }
-                    return;
-                  } catch (writeErr) {
-                    process.stderr.write(`[tee] salvage write failed: ${writeErr.message ?? writeErr}
-`);
-                  }
-                }
-                try {
-                  unlinkSync2(saveToFile);
-                } catch {
-                }
-                process.stdout.write(theme.error(
-                  `
-  \u2717 Rejected save: response was pseudo-tool-call markup with no usable document body (matched: ${pseudoMatch})
-  ${saveToFile} was deleted; asking model to retry.
-
-`
-                ));
-                const errorResults = result.toolCalls.map((tc) => ({
-                  callId: tc.id,
-                  content: tc.name === "save_last_response" ? `[save_last_response REJECTED] Your output was tool-call XML/JSON with no document body. ${saveToFile} was NOT saved.
-
-This fresh stream has NO tools \u2014 output is captured verbatim. STOP emitting <tool_call>, <function_calls>, <invoke>, <think>, or JSON tool blocks. Produce the document body NOW: start with a markdown heading like "# \u5BA1\u8BA1\u62A5\u544A" and write the full report.` : `[skipped: save_last_response was rejected and other parallel calls are abandoned]`,
-                  isError: tc.name === "save_last_response"
-                }));
-                const reasoningContent3 = "reasoningContent" in result ? result.reasoningContent : void 0;
-                const newMsgs3 = provider.buildToolResultMessages(result.toolCalls, errorResults, reasoningContent3);
-                extraMessages.push(...newMsgs3);
-                if (genUsage) {
-                  roundUsage.inputTokens += genUsage.inputTokens;
-                  roundUsage.outputTokens += genUsage.outputTokens;
-                  roundUsage.cacheCreationTokens += genUsage.cacheCreationTokens ?? 0;
-                  roundUsage.cacheReadTokens += genUsage.cacheReadTokens ?? 0;
-                }
-                continue;
-              }
-              lastResponseStore.content = genContent;
-              if (genUsage) {
-                roundUsage.inputTokens += genUsage.inputTokens;
-                roundUsage.outputTokens += genUsage.outputTokens;
-                roundUsage.cacheCreationTokens += genUsage.cacheCreationTokens ?? 0;
-                roundUsage.cacheReadTokens += genUsage.cacheReadTokens ?? 0;
-              }
-              session.addMessage({ role: "assistant", content: genContent, timestamp: /* @__PURE__ */ new Date() });
-              this.events.emit("message.after", { content: genContent });
-              const lines = genContent.split("\n").length;
-              const bytes = Buffer.byteLength(genContent, "utf-8");
-              const syntheticResults = result.toolCalls.map((tc) => ({
-                callId: tc.id,
-                content: tc.name === "save_last_response" ? `File saved: ${saveToFile} (${lines} lines, ${bytes} bytes)` : `[skipped: file already saved by tee streaming]`,
-                isError: false
-              }));
-              const reasoningContent2 = "reasoningContent" in result ? result.reasoningContent : void 0;
-              const newMsgs2 = provider.buildToolResultMessages(result.toolCalls, syntheticResults, reasoningContent2);
-              extraMessages.push(...newMsgs2);
-              if (roundUsage.inputTokens > 0 || roundUsage.outputTokens > 0) {
-                this.addSessionUsage(roundUsage, effectiveModel);
-                session.addTokenUsage(roundUsage);
-                if (teeShowTokens && !teeTokShown) {
-                  this.renderer.renderUsage(roundUsage, this.sessionTokenUsage);
-                }
+          this.events.emit("message.after", { content });
+        },
+        onInterrupted: () => {
+          spinner.stop();
+          process.stdout.write(theme.warning("\n\u26A1 Interrupted by user (Ctrl+C) \u2014 agentic loop stopped.\n"));
+        },
+        persistRound: (toolCalls, results, info) => {
+          persistToolRound(session, toolCalls, results, info);
+        },
+        onMcpToolUsed: (name) => usedMcpToolNames.add(name),
+        onToolsExecuted: (toolCalls, _results, extraMessages) => {
+          const readFileCalls = toolCalls.filter((tc) => tc.name === "read_file");
+          for (const rfc of readFileCalls) {
+            const filePath = rfc.arguments?.path;
+            if (filePath) {
+              const fileReadCount = extraMessages.filter((msg) => {
+                const m = msg;
+                if (m.role !== "assistant") return false;
+                const tcs = m.tool_calls;
+                if (!Array.isArray(tcs)) return false;
+                return tcs.some((tc) => {
+                  const fn = tc.function;
+                  return fn?.name === "read_file" && JSON.stringify(fn?.arguments ?? "").includes(filePath);
+                });
+              }).length;
+              if (fileReadCount >= 2) {
+                extraMessages.push({
+                  role: "user",
+                  content: `\u26A0\uFE0F You have read the file "${filePath}" ${fileReadCount + 1} times already. The content hasn't changed \u2014 do NOT read it again. Use the information you already have.`
+                });
               }
-            } finally {
-              this.teardownStreamInterrupt();
             }
-            return;
           }
-        }
-        askUserContext.rl = this.rl;
-        googleSearchContext.configManager = this.config;
-        streamToFileContext.provider = provider;
-        streamToFileContext.model = effectiveModel;
-        streamToFileContext.systemPrompt = systemPromptVolatile ? `${systemPrompt}
-
----
-
-${systemPromptVolatile}` : systemPrompt;
-        streamToFileContext.messages = apiMessages;
-        streamToFileContext.extraMessages = extraMessages;
-        streamToFileContext.temperature = modelParams.temperature;
-        streamToFileContext.timeout = modelParams.timeout;
-        spawnAgentContext.provider = provider;
-        spawnAgentContext.model = effectiveModel;
-        spawnAgentContext.systemPrompt = systemPromptVolatile ? `${systemPrompt}
-
----
-
-${systemPromptVolatile}` : systemPrompt;
-        spawnAgentContext.modelParams = modelParams;
-        spawnAgentContext.configManager = this.config;
-        ToolExecutor.currentMessageIndex = session.messages.length;
-        const toolResults = await this.toolExecutor.executeAll(result.toolCalls);
-        if (isInterrupted()) {
-          spinner.stop();
-          process.stdout.write(theme.warning("\n\u26A1 Interrupted by user (Ctrl+C) \u2014 agentic loop stopped.\n"));
-          this.teardownInterjectionListener();
-          return;
-        }
-        const thisRoundTools = result.toolCalls.map((tc) => tc.name);
-        roundToolHistory.push({ round: round + 1, tools: thisRoundTools });
-        const readFileCalls = result.toolCalls.filter((tc) => tc.name === "read_file");
-        for (const rfc of readFileCalls) {
-          const filePath = rfc.arguments?.path;
-          if (filePath) {
-            const fileReadCount = extraMessages.filter((msg) => {
-              const m = msg;
-              if (m.role !== "assistant") return false;
-              const tcs = m.tool_calls;
-              if (!Array.isArray(tcs)) return false;
-              return tcs.some((tc) => {
-                const fn = tc.function;
-                return fn?.name === "read_file" && JSON.stringify(fn?.arguments ?? "").includes(filePath);
-              });
-            }).length;
-            if (fileReadCount >= 2) {
+          process.stdin.resume();
+        },
+        afterToolRoundPersist: (toolCalls, _results, extraMessages) => {
+          const thisRoundHadWrite = toolCalls.some(
+            (tc) => tc.name === "write_file" || tc.name === "edit_file"
+          );
+          if (thisRoundHadWrite) {
+            const totalWritten = extractWrittenFilePaths(extraMessages).length;
+            if (totalWritten > 0) {
               extraMessages.push({
                 role: "user",
-                content: `\u26A0\uFE0F You have read the file "${filePath}" ${fileReadCount + 1} times already. The content hasn't changed \u2014 do NOT read it again. Use the information you already have.`
+                content: buildWriteRoundReminder(totalWritten)
               });
             }
           }
-        }
-        process.stdin.resume();
-        const reasoningContent = "reasoningContent" in result ? result.reasoningContent : void 0;
-        const newMsgs = provider.buildToolResultMessages(result.toolCalls, toolResults, reasoningContent);
-        extraMessages.push(...newMsgs);
-        for (const tc of result.toolCalls) {
-          if (tc.name.startsWith("mcp__")) usedMcpToolNames.add(tc.name);
-        }
-        const streamedContent = "content" in result ? result.content : void 0;
-        persistToolRound(session, result.toolCalls, toolResults, {
-          assistantContent: streamedContent,
-          reasoningContent
-        });
-        const thisRoundHadWrite = result.toolCalls.some(
-          (tc) => tc.name === "write_file" || tc.name === "edit_file"
-        );
-        if (thisRoundHadWrite) {
-          const totalWritten = extractWrittenFilePaths(extraMessages).length;
-          if (totalWritten > 0) {
-            extraMessages.push({
-              role: "user",
-              content: buildWriteRoundReminder(totalWritten)
-            });
-          }
-        }
-        if (freeRounds.apply(result.toolCalls.map((tc) => tc.name))) {
-          round--;
-        }
-        const currentSignature = result.toolCalls.map((tc) => `${tc.name}:${JSON.stringify(tc.arguments)}`).join("|");
-        if (currentSignature === lastToolCallSignature) {
-          repeatedToolCallCount++;
-          if (repeatedToolCallCount >= MAX_REPEATED_TOOL_CALLS) {
-            spinner.stop();
-            process.stderr.write(
-              theme.warning(`
+        },
+        checkLoopHealth: (toolCalls, extraMessages) => {
+          const currentSignature = toolCalls.map((tc) => `${tc.name}:${JSON.stringify(tc.arguments)}`).join("|");
+          if (currentSignature === lastToolCallSignature) {
+            repeatedToolCallCount++;
+            if (repeatedToolCallCount >= MAX_REPEATED_TOOL_CALLS) {
+              spinner.stop();
+              process.stderr.write(
+                theme.warning(`
 \u26A0 Detected ${repeatedToolCallCount + 1} identical consecutive tool calls \u2014 breaking loop.
 `)
-            );
-            extraMessages.push({
-              role: "user",
-              content: "You are stuck in a loop calling the same tool with the same arguments repeatedly. Stop calling tools and give the user a final text response summarizing what you found and what needs to be done next. " + (this.planMode ? "If you need to execute commands (bash, psql, etc.), tell the user to type `/plan execute` first." : "")
-            });
+              );
+              extraMessages.push({
+                role: "user",
+                content: "You are stuck in a loop calling the same tool with the same arguments repeatedly. Stop calling tools and give the user a final text response summarizing what you found and what needs to be done next. " + (this.planMode ? "If you need to execute commands (bash, psql, etc.), tell the user to type `/plan execute` first." : "")
+              });
+              repeatedToolCallCount = 0;
+              lastToolCallSignature = "";
+              return "skip-checkpoint";
+            }
+          } else {
+            lastToolCallSignature = currentSignature;
             repeatedToolCallCount = 0;
-            lastToolCallSignature = "";
-            continue;
           }
-        } else {
-          lastToolCallSignature = currentSignature;
-          repeatedToolCallCount = 0;
-        }
-        if (this._userInterjection) {
-          const msg = this._userInterjection;
-          this._userInterjection = null;
-          process.stdout.write(theme.warning(`\u26A1 Interjection: "${msg}"
-`));
-          extraMessages.push({ role: "user", content: msg });
-        }
-        const effectiveRound = round + 1;
-        const remaining = maxToolRounds - effectiveRound;
-        if (autoPauseInterval > 0 && effectiveRound > 0 && effectiveRound % autoPauseInterval === 0 && remaining > 0) {
+        },
+        requestAutoPause: async ({ effectiveRound, maxToolRounds: totalRounds, remaining, toolSummary }) => {
           spinner.stop();
           process.stdout.write("\n");
-          process.stdout.write(theme.warning(`\u23F8  Auto-pause: ${effectiveRound}/${maxToolRounds} rounds used, ${remaining} remaining
+          process.stdout.write(theme.warning(`\u23F8  Auto-pause: ${effectiveRound}/${totalRounds} rounds used, ${remaining} remaining
 `));
-          const summary = summarizeRecentTools(roundToolHistory, autoPauseInterval);
-          if (summary) {
-            process.stdout.write(theme.dim(`  Tools used: ${summary}
+          if (toolSummary) {
+            process.stdout.write(theme.dim(`  Tools used: ${toolSummary}
 `));
           }
           process.stdout.write(theme.dim("  Press ") + theme.info("y") + theme.dim(" to continue, or ") + theme.info("type a message") + theme.dim(" to redirect AI:\n"));
@@ -6988,74 +6736,222 @@ ${systemPromptVolatile}` : systemPrompt;
           this.setupInterjectionListener();
           if (pauseResponse === "n" || pauseResponse === "N" || pauseResponse === "\x1B") {
             process.stdout.write(theme.warning("\u26A1 Stopped by user at auto-pause checkpoint\n"));
-            extraMessages.push({ role: "user", content: buildUserStopMessage(effectiveRound, maxToolRounds) });
-            break;
-          } else if (pauseResponse && pauseResponse !== "y" && pauseResponse !== "Y" && pauseResponse !== "") {
+            return { action: "stop" };
+          }
+          if (pauseResponse && pauseResponse !== "y" && pauseResponse !== "Y") {
             process.stdout.write(theme.warning(`\u26A1 Redirect: "${pauseResponse}"
 `));
-            extraMessages.push({ role: "user", content: pauseResponse });
+            process.stdout.write(theme.success(`\u25B6 Continuing... (${remaining} rounds left)
+`));
+            return { action: "redirect", message: pauseResponse };
           }
           process.stdout.write(theme.success(`\u25B6 Continuing... (${remaining} rounds left)
 `));
-        }
-        const nextRound = round + 2;
-        spinner.start(
-          nextRound <= maxToolRounds ? `Thinking... (round ${nextRound}/${maxToolRounds})` : "Thinking..."
-        );
-      }
-      spinner.stop();
-      try {
-        spinner.start("Generating summary...");
-        const summaryExtra = [
-          ...extraMessages,
-          { role: "user", content: buildRoundsExhaustedPrompt(maxToolRounds) }
-        ];
-        const summaryResult = await provider.chatWithTools(
-          {
-            messages: apiMessages,
-            model: effectiveModel,
-            systemPrompt,
-            systemPromptVolatile,
-            stream: false,
-            temperature: modelParams.temperature,
-            maxTokens: modelParams.maxTokens,
-            timeout: modelParams.timeout,
-            thinking: modelParams.thinking,
-            thinkingBudget: modelParams.thinkingBudget,
-            _extraMessages: summaryExtra
-          },
-          []
-          // 不提供任何工具，强制 AI 返回纯文本
-        );
-        spinner.stop();
-        if ("content" in summaryResult) {
-          this.renderer.renderError(`Reached maximum tool call rounds (${maxToolRounds}). Here is a summary:`);
-          this.renderer.renderResponse(summaryResult.content);
-          lastResponseStore.content = summaryResult.content;
-          session.addMessage({ role: "assistant", content: summaryResult.content, timestamp: /* @__PURE__ */ new Date() });
-          if (summaryResult.usage) {
-            roundUsage.inputTokens += summaryResult.usage.inputTokens;
-            roundUsage.outputTokens += summaryResult.usage.outputTokens;
-            roundUsage.cacheCreationTokens += summaryResult.usage.cacheCreationTokens ?? 0;
-            roundUsage.cacheReadTokens += summaryResult.usage.cacheReadTokens ?? 0;
-          }
-        } else {
-          this.renderer.renderError(
-            `Reached maximum tool call rounds (${maxToolRounds}). Stopping.
-Tip: You can continue the conversation by asking the AI to proceed.`
+          return { action: "continue" };
+        },
+        onRoundEnd: (round) => {
+          const nextRound = round + 2;
+          spinner.start(
+            nextRound <= maxToolRounds ? `Thinking... (round ${nextRound}/${maxToolRounds})` : "Thinking..."
           );
-        }
-      } catch {
-        this.renderer.renderError(
-          `Reached maximum tool call rounds (${maxToolRounds}). Stopping.
+        },
+        onRoundsExhausted: (summaryContent) => {
+          if (summaryContent !== null) {
+            this.renderer.renderError(`Reached maximum tool call rounds (${maxToolRounds}). Here is a summary:`);
+            this.renderer.renderResponse(summaryContent);
+            lastResponseStore.content = summaryContent;
+            session.addMessage({ role: "assistant", content: summaryContent, timestamp: /* @__PURE__ */ new Date() });
+          } else {
+            this.renderer.renderError(
+              `Reached maximum tool call rounds (${maxToolRounds}). Stopping.
 Tip: You can continue the conversation by asking the AI to proceed.`
-        );
-      }
-      if (roundUsage.inputTokens > 0 || roundUsage.outputTokens > 0) {
-        this.addSessionUsage(roundUsage, effectiveModel);
-        session.addTokenUsage(roundUsage);
+            );
+          }
+        },
+        // ─── save_last_response 特殊处理（tee 流式写盘）────────────────────────
+        // 架构设计：
+        //   AI 在工具调用轮次中调用 save_last_response(path)，意图是"生成内容 + 保存"。
+        //   旧方案：AI 先输出内容 → lastResponseStore 暂存 → 工具读 store 写盘
+        //     问题：Kimi API 会截断 tool_call arguments (~2KB)，内容无法通过参数传递；
+        //           第一轮调用时 store 里只有上一轮旧内容，根本没有本次生成的内容。
+        //   新方案（tee streaming）：
+        //     1. core 检测到 save_last_response 调用并从参数中提取目标文件路径
+        //     2. 此钩子发起流式请求，renderStream 同时写入终端 + 磁盘（tee 模式）
+        //     3. 注入合成的工具成功结果到 extraMessages（跳过实际工具执行）
+        //     4. 成功返回 'stop' 直接结束（usage 已在钩子内入账）；失败/拒绝返回
+        //        'continue' 让模型重试
+        runSaveLastResponseTee: async ({ toolCalls, saveToFile, extraMessages, reasoningContent }) => {
+          spinner.stop();
+          const teeAc = this.setupStreamInterrupt();
+          try {
+            const teeSystemPrompt = stripToolCallReminder(systemPrompt ?? "") + CONTENT_ONLY_STREAM_REMINDER;
+            const teeExtraMessages = extraMessages.length > 0 ? [...extraMessages, { role: "user", content: TEE_FINAL_USER_NUDGE }] : [{ role: "user", content: TEE_FINAL_USER_NUDGE }];
+            const genStream = provider.chatStream({
+              messages: apiMessages,
+              model: effectiveModel,
+              systemPrompt: teeSystemPrompt,
+              systemPromptVolatile,
+              stream: true,
+              temperature: modelParams.temperature,
+              maxTokens: modelParams.maxTokens,
+              timeout: modelParams.timeout,
+              thinking: modelParams.thinking,
+              thinkingBudget: modelParams.thinkingBudget,
+              signal: teeAc.signal,
+              _extraMessages: teeExtraMessages
+            });
+            const teeShowTokens = this.shouldShowTokens();
+            let genContent;
+            let genUsage;
+            let teeTokShown = false;
+            try {
+              const teeResult = await this.renderer.renderStream(
+                genStream,
+                { saveToFile, showTokens: teeShowTokens, sessionTotal: teeShowTokens ? { ...this.sessionTokenUsage } : void 0, signal: teeAc.signal }
+              );
+              genContent = teeResult.content;
+              genUsage = teeResult.usage;
+              teeTokShown = teeResult.tokensShown;
+            } catch (teeErr) {
+              try {
+                unlinkSync2(saveToFile);
+              } catch {
+              }
+              const errMsg = teeErr instanceof Error ? teeErr.message : String(teeErr);
+              process.stdout.write(theme.error(
+                `
+  \u2717 tee stream failed: ${errMsg}
+  ${saveToFile} (partial) was deleted. Asking model to retry.
+
+`
+              ));
+              const errorResults = toolCalls.map((tc) => ({
+                callId: tc.id,
+                content: tc.name === "save_last_response" ? `[save_last_response failed] streaming was interrupted: ${errMsg}. ${saveToFile} was NOT saved. Retry \u2014 and consider producing a more compact output (split very large reports across multiple save_last_response calls if the previous attempt timed out).` : `[skipped: save_last_response failed]`,
+                isError: tc.name === "save_last_response"
+              }));
+              const newMsgs2 = provider.buildToolResultMessages(toolCalls, errorResults, reasoningContent);
+              extraMessages.push(...newMsgs2);
+              return "continue";
+            }
+            const metaMatch = detectMetaNarration(genContent);
+            if (metaMatch) {
+              try {
+                unlinkSync2(saveToFile);
+              } catch {
+              }
+              process.stdout.write(theme.error(
+                `
+  \u2717 Rejected save: response was meta-narration / leaked reasoning, not document body (matched: ${metaMatch})
+  ${saveToFile} was deleted; asking model to retry.
+
+`
+              ));
+              const errorResults = toolCalls.map((tc) => ({
+                callId: tc.id,
+                content: tc.name === "save_last_response" ? `[save_last_response REJECTED] Your output was internal reasoning / meta-narration about the task (e.g. "Let me re-read\u2026", "the user is asking me to\u2026") instead of the requested document body. ${saveToFile} was NOT saved.
+
+This fresh stream has NO tools. Produce ONLY the document body: start with a markdown heading like "# \u5BA1\u8BA1\u62A5\u544A" / "# Audit Report" and write the full content. Do NOT narrate that you will produce the document \u2014 produce it.` : `[skipped: save_last_response was rejected and other parallel calls are abandoned]`,
+                isError: tc.name === "save_last_response"
+              }));
+              const newMsgs2 = provider.buildToolResultMessages(toolCalls, errorResults, reasoningContent);
+              extraMessages.push(...newMsgs2);
+              if (genUsage) accumulateUsage(usage, genUsage);
+              return "continue";
+            }
+            const pseudoMatch = detectPseudoToolCalls(genContent);
+            if (pseudoMatch) {
+              const cleaned = stripPseudoToolCalls(genContent);
+              if (looksLikeDocumentBody(cleaned)) {
+                try {
+                  writeFileSync2(saveToFile, cleaned, "utf-8");
+                  process.stdout.write(theme.warning(
+                    `
+  \u26A0 Salvaged save: stripped pseudo-tool-call markup (matched: ${pseudoMatch})
+  ${saveToFile} now contains the cleaned document (${cleaned.length} chars; was ${genContent.length}).
+
+`
+                  ));
+                  lastResponseStore.content = cleaned;
+                  if (genUsage) accumulateUsage(usage, genUsage);
+                  session.addMessage({ role: "assistant", content: cleaned, timestamp: /* @__PURE__ */ new Date() });
+                  this.events.emit("message.after", { content: cleaned });
+                  const lines2 = cleaned.split("\n").length;
+                  const bytes2 = Buffer.byteLength(cleaned, "utf-8");
+                  const okResults = toolCalls.map((tc) => ({
+                    callId: tc.id,
+                    content: tc.name === "save_last_response" ? `File saved (with cleanup): ${saveToFile} (${lines2} lines, ${bytes2} bytes; pseudo-tool-call markup was stripped before save)` : `[skipped: file already saved by tee streaming]`,
+                    isError: false
+                  }));
+                  const newMsgs3 = provider.buildToolResultMessages(toolCalls, okResults, reasoningContent);
+                  extraMessages.push(...newMsgs3);
+                  if (usage.inputTokens > 0 || usage.outputTokens > 0) {
+                    this.addSessionUsage(usage, effectiveModel);
+                    session.addTokenUsage(usage);
+                    if (teeShowTokens && !teeTokShown) {
+                      this.renderer.renderUsage(usage, this.sessionTokenUsage);
+                    }
+                  }
+                  return "stop";
+                } catch (writeErr) {
+                  process.stderr.write(`[tee] salvage write failed: ${writeErr.message ?? writeErr}
+`);
+                }
+              }
+              try {
+                unlinkSync2(saveToFile);
+              } catch {
+              }
+              process.stdout.write(theme.error(
+                `
+  \u2717 Rejected save: response was pseudo-tool-call markup with no usable document body (matched: ${pseudoMatch})
+  ${saveToFile} was deleted; asking model to retry.
+
+`
+              ));
+              const errorResults = toolCalls.map((tc) => ({
+                callId: tc.id,
+                content: tc.name === "save_last_response" ? `[save_last_response REJECTED] Your output was tool-call XML/JSON with no document body. ${saveToFile} was NOT saved.
+
+This fresh stream has NO tools \u2014 output is captured verbatim. STOP emitting <tool_call>, <function_calls>, <invoke>, <think>, or JSON tool blocks. Produce the document body NOW: start with a markdown heading like "# \u5BA1\u8BA1\u62A5\u544A" and write the full report.` : `[skipped: save_last_response was rejected and other parallel calls are abandoned]`,
+                isError: tc.name === "save_last_response"
+              }));
+              const newMsgs2 = provider.buildToolResultMessages(toolCalls, errorResults, reasoningContent);
+              extraMessages.push(...newMsgs2);
+              if (genUsage) accumulateUsage(usage, genUsage);
+              return "continue";
+            }
+            lastResponseStore.content = genContent;
+            if (genUsage) accumulateUsage(usage, genUsage);
+            session.addMessage({ role: "assistant", content: genContent, timestamp: /* @__PURE__ */ new Date() });
+            this.events.emit("message.after", { content: genContent });
+            const lines = genContent.split("\n").length;
+            const bytes = Buffer.byteLength(genContent, "utf-8");
+            const syntheticResults = toolCalls.map((tc) => ({
+              callId: tc.id,
+              content: tc.name === "save_last_response" ? `File saved: ${saveToFile} (${lines} lines, ${bytes} bytes)` : `[skipped: file already saved by tee streaming]`,
+              isError: false
+            }));
+            const newMsgs = provider.buildToolResultMessages(toolCalls, syntheticResults, reasoningContent);
+            extraMessages.push(...newMsgs);
+            if (usage.inputTokens > 0 || usage.outputTokens > 0) {
+              this.addSessionUsage(usage, effectiveModel);
+              session.addTokenUsage(usage);
+              if (teeShowTokens && !teeTokShown) {
+                this.renderer.renderUsage(usage, this.sessionTokenUsage);
+              }
+            }
+            return "stop";
+          } finally {
+            this.teardownStreamInterrupt();
+          }
+        }
+      });
+      if (loopResult.reason !== "tee-stop" && (usage.inputTokens > 0 || usage.outputTokens > 0)) {
+        this.addSessionUsage(usage, effectiveModel);
+        session.addTokenUsage(usage);
         if (this.shouldShowTokens()) {
-          this.renderer.renderUsage(roundUsage, this.sessionTokenUsage);
+          this.renderer.renderUsage(usage, this.sessionTokenUsage);
         }
       }
     } finally {
@@ -7360,7 +7256,7 @@ program.command("web").description("Start Web UI server with browser-based chat
     console.error("Error: Invalid port number. Must be between 1 and 65535.");
     process.exit(1);
   }
-  const { startWebServer } = await import("./server-UT6PLLZC.js");
+  const { startWebServer } = await import("./server-2B5JDVJS.js");
   await startWebServer({ port, host: options.host });
 });
 program.command("user [action] [username]").description("Manage Web UI users (list | create <name> | delete <name> | reset-password <name> | logout-all <name> | migrate <name>)").action(async (action, username) => {
@@ -7527,16 +7423,16 @@ program.command("sessions").description("List recent conversation sessions").opt
   console.log(footer + "\n");
 });
 program.command("usage").description("Show token + cost usage grouped by provider/model (cross-session)").option("--days <n>", "Only the last N days (inclusive of today)").option("--month <ym>", "Only a specific month, format YYYY-MM (e.g. 2026-06)").option("--json", "Output as JSON (for scripting)").action(async (options) => {
-  const { runUsageCli } = await import("./usage-5KBD4UBB.js");
+  const { runUsageCli } = await import("./usage-ZVKFH7BM.js");
   await runUsageCli(options);
 });
 program.command("doctor").description("Health check: API keys, config, MCP, recent crashes, tool usage, disk usage").option("--json", "Output as JSON (for scripting)").option("--reset-stats", "Reset accumulated tool usage statistics").action(async (options) => {
-  const { runDoctorCli } = await import("./doctor-cli-X6MOE3YE.js");
+  const { runDoctorCli } = await import("./doctor-cli-MYJFAWKV.js");
   await runDoctorCli({ json: !!options.json, resetStats: !!options.resetStats });
 });
 program.command("batch <action> [arg] [arg2]").description("Anthropic Message Batches: submit | list | status <id> | results <id> [out] | cancel <id>").option("--dry-run", "Parse and validate input without submitting (submit only)").action(async (action, arg, arg2, options) => {
   try {
-    const batch = await import("./batch-ILD2EPEO.js");
+    const batch = await import("./batch-UTP6NYVX.js");
     switch (action) {
       case "submit":
         if (!arg) {
@@ -7579,7 +7475,7 @@ program.command("batch <action> [arg] [arg2]").description("Anthropic Message Ba
   }
 });
 program.command("mcp-serve").description("Start an MCP server over STDIO, exposing aicli's built-in tools to Claude Desktop / Cursor / other MCP clients").option("--allow-destructive", "Allow bash / run_interactive / task_create (always destructive in MCP mode)").option("--allow-outside-cwd", "Allow tool path arguments to escape the sandbox root \u2014 disabled by default").option("--tools <list>", "Comma-separated whitelist of tools to expose (default: all eligible tools)").option("--cwd <path>", "Working directory AND sandbox root (default: current directory)").action(async (options) => {
-  const { startMcpServer } = await import("./server-H3KIFOLK.js");
+  const { startMcpServer } = await import("./server-25WVH5YX.js");
   await startMcpServer({
     allowDestructive: !!options.allowDestructive,
     allowOutsideCwd: !!options.allowOutsideCwd,
@@ -7588,7 +7484,7 @@ program.command("mcp-serve").description("Start an MCP server over STDIO, exposi
   });
 });
 program.command("ci").description("Headless PR review (code + security) \u2014 reads git/gh diff, optionally posts to PR. Designed for GitHub Actions.").option("--pr <num>", "PR number; diff fetched via `gh pr diff <num>`", (v) => parseInt(v, 10)).option("--base <ref>", "Base ref for `git diff <ref>...HEAD` (ignored when --pr set)").option("--post", "Post review as a PR comment (requires gh CLI + GH_TOKEN, needs --pr)").option("--no-update", "Always create a new comment instead of updating the previous aicli review").option("--skip-code", "Skip the code review section").option("--skip-security", "Skip the security review section").option("--detailed", "Use the detailed code-review prompt").option("--max-diff <n>", "Max diff chars sent to the model (default 30000)", (v) => parseInt(v, 10)).option("--provider <id>", "Override provider (default: config.defaultProvider)").option("--model <id>", "Override model").option("--dry-run", "Print result to stdout instead of posting (overrides --post)").action(async (options) => {
-  const { runCi } = await import("./ci-7YWXFKGE.js");
+  const { runCi } = await import("./ci-2WFKSG2J.js");
   const result = await runCi({
     pr: options.pr,
     base: options.base,
@@ -7734,7 +7630,7 @@ program.command("hub [topic]").description("Start multi-agent hub (discuss / bra
     }),
     config.get("customProviders")
   );
-  const { startHub } = await import("./hub-SFMWUEUW.js");
+  const { startHub } = await import("./hub-CHE7JDIH.js");
   await startHub(
     {
       topic: topic ?? "",