Method run

src/main/java/com/tcode/agent/Agent.java:119–250

运行 Agent 循环

(String userInput)

     * 运行 Agent 循环
     */
    public String run(String userInput) {
        log.info("ReAct run started: inputLength={}", userInput == null ? 0 : userInput.length());
        pruneHistoricalImagePayloads();
        // 存入短期记忆
        memoryManager.addUserMessage(userInput);
        storeExplicitBrowserMemoryHint(userInput);

        // 检索相关长期记忆，注入到 system prompt
        ContextProfile contextProfile = memoryManager.getContextProfile();
        String memoryContext = memoryManager.buildContextForQuery(userInput, contextProfile.memoryContextTokens());
        updateSystemPromptWithMemory(memoryContext);

        // 添加用户输入到历史（如有 skill body 注入，前置到原文之前）
        String userMessageContent = prependSkillBodies(userInput);
        conversationHistory.add(ImageReferenceParser.userMessage(
                userMessageContent,
                Path.of(toolRegistry.getProjectPath())));
        StringBuilder reasoningTranscript = new StringBuilder();
        StreamRenderer streamRenderer = new StreamRenderer(renderer());

        long startNanos = System.nanoTime();
        AgentBudget budget = AgentBudget.fromLlmClient(llmClient);
        pushStatus(budget, startNanos, "running");

        // 主退出条件 = LLM 自己决定（不再调用工具就返回）；
        // budget 仅在 token 用尽 / 检测到死循环 / 超出硬轮数时兜底。
        while (true) {
            if (CancellationContext.isCancelled()) {
                log.info("ReAct run cancelled before iteration");
                pushStatus(budget, startNanos, "idle");
                return "⏹️ 已取消当前任务。";
            }
            // 调 LLM 前评估 conversationHistory 是否接近 window 上限；超阈值就把早期消息压缩成摘要。
            // 这是与第 3 期 Memory 短期记忆压缩并行的另一道压缩——后者只压 shortTermMemory，
            // 真正决定下一轮 LLM input token 的是这里。
            injectPendingLspDiagnostics();
            maybeCompactHistory();
            AgentBudget.ExitReason exitReason = budget.check();
            if (exitReason != AgentBudget.ExitReason.WITHIN_BUDGET) {
                String description = budget.describeExit(exitReason);
                log.warn("ReAct run exhausted budget: reason={}, iteration={}, tokens={}/{}",
                        exitReason, budget.iteration(),
                        budget.totalInputTokens() + budget.totalOutputTokens(), budget.tokenBudget());
                pushStatus(budget, startNanos, "idle");
                return "❌ " + description;
            }

            int iteration = budget.beginIteration();

            try {
                List<LlmClient.Tool> toolDefinitions = toolRegistry.getToolDefinitions();
                logRequestContext("react iteration=" + iteration, toolDefinitions);
                streamRenderer.beginThinking();
                // 调用 LLM
                LlmClient.ChatResponse response = llmClient.chat(
                        conversationHistory,
                        toolDefinitions,
                        streamRenderer