fix: only load project context (CODEACTOR.md/CLAUDE.md) on first conv…

README.md

@@ -185,39 +185,6 @@ Disable Meta-Agent via startup flag:
 ./codeactor tui --disable-agents=meta
 ```
 
-## DevOps-Agent
-
-The **DevOps-Agent** is the operations and infrastructure specialist — it handles all non-coding operational tasks by executing shell commands, inspecting the file system, and analyzing command output. When the Conductor encounters a task involving system administration, log inspection, process management, or ad-hoc shell commands, it delegates to the DevOps-Agent via `delegate_devops`.
-
-### Capabilities
-
-- **Shell Command Execution** (`run_bash`) — Run any bash command with foreground/background support, danger detection, and workspace boundary checks
-- **File System Inspection** — `read_file`, `list_dir`, `print_dir_tree`, `search_by_regex` for browsing logs, configs, and directories
-- **Self-Correction** — `thinking` tool for analyzing command failures and adjusting approach before retrying
-- **Isolated Analysis** — `micro_agent` for deep reasoning on command output or generating structured reports
-
-### Example use cases
-
-- Check disk usage, memory, and system resources
-- Find all log files modified in the last 24 hours
-- Restart services or check process status
-- Inspect configuration files
-- Run system diagnostics and generate reports
-- Execute ad-hoc shell pipelines for data processing
-
-### Configuration
-
-```toml
-[agent]
-devops_max_steps = 15    # Max LLM steps for DevOps-Agent (default: 15)
-```
-
-Disable DevOps-Agent via startup flag:
-
-```bash
-./codeactor tui --disable-agents=devops
-```
-
 ## Codebase Analysis Engine
 
 The `codeactor-codebase` is a standalone **Rust** service that provides deep code analysis capabilities. It runs as a background HTTP server managed automatically by the Go binary.

README_zh.md

@@ -184,39 +184,6 @@ meta_retry_count = 5   # JSON 解析失败重试次数（默认 5）
 ./codeactor tui --disable-agents=meta
 ```
 
-## DevOps-Agent（运维代理）
-
-**DevOps-Agent** 是运维和基础设施专家——通过执行 Shell 命令、检查文件系统和分析命令输出来处理所有非编码的运维任务。当 Conductor 遇到系统管理、日志检查、进程管理或 ad-hoc shell 命令类任务时，会通过 `delegate_devops` 委派给 DevOps-Agent。
-
-### 核心能力
-
-- **Shell 命令执行** (`run_bash`) — 运行任意 bash 命令，支持前台/后台运行，含危险检测和工作空间边界检查
-- **文件系统检查** — `read_file`、`list_dir`、`print_dir_tree`、`search_by_regex` 用于浏览日志、配置和目录
-- **自我修正** — 使用 `thinking` 工具分析命令失败原因，调整策略后重试
-- **独立分析** — 使用 `micro_agent` 对命令输出进行深度推理或生成结构化报告
-
-### 示例用例
-
-- 检查磁盘使用率、内存和系统资源
-- 查找最近 24 小时内修改的所有日志文件
-- 重启服务或检查进程状态
-- 检查配置文件
-- 运行系统诊断并生成报告
-- 执行 ad-hoc shell 管道进行数据处理
-
-### 配置
-
-```toml
-[agent]
-devops_max_steps = 15    # DevOps-Agent 最大 LLM 步数（默认 15）
-```
-
-通过启动参数禁用 DevOps-Agent：
-
-```bash
-./codeactor tui --disable-agents=devops
-```
-
 ## Codebase 分析引擎
 
 `codeactor-codebase` 是一个独立的 **Rust** 服务，提供深度代码分析能力。它作为后台 HTTP 服务器运行，由 Go 二进制自动管理。

codebase/src/http/handlers/mod.rs

@@ -871,6 +871,15 @@ pub(crate) fn setup_watcher(
         match res {
             Ok(event) => {
                 if event.kind.is_modify() || event.kind.is_create() || event.kind.is_remove() {
+                    // Filter out events from .git directory
+                    let should_process = event.paths.iter().any(|path| {
+                        !path.components().any(|comp| comp.as_os_str() == ".git")
+                    });
+
+                    if !should_process {
+                        return;
+                    }
+
                      tracing::info!("File change detected: {:?}, queueing re-analysis", event.paths);
                      let _ = tx.send(());
                 }

internal/agents/conductor.go

@@ -67,12 +67,17 @@ type ConductorAgent struct {
 	compactEngine  *compact.Engine           // 上下文压缩引擎
 	compactConfig  *compact.Config           // 压缩配置
 	summaryEngine  llm.Engine                // 独立的摘要 LLM 引擎（nil 则复用主引擎）
+	cachedProjectContext *ProjectContextLoadResult // 缓存项目上下文文件（同一会话只加载一次）
 }
 
 // loadProjectContext 读取工作区目录下的项目上下文文件（CODEACTOR.md、CLAUDE.md、AGENTS.md），
 // 将成功读取的文件内容格式化后组合返回。文件按顺序尝试，不存在或读取失败时忽略。
 // 返回加载的文件列表和组合后的内容。
 func (a *ConductorAgent) loadProjectContext() *ProjectContextLoadResult {
+	// 如果已经加载过，直接返回缓存（同一 Agent 实例会话内只加载一次）
+	if a.cachedProjectContext != nil {
+		return a.cachedProjectContext
+	}
 	result := &ProjectContextLoadResult{
 		LoadedFiles: []ProjectContextFile{},
 	}
@@ -95,6 +100,8 @@ func (a *ConductorAgent) loadProjectContext() *ProjectContextLoadResult {
 		}
 	}
 	result.Content = sb.String()
+	// 缓存加载结果，避免后续调用重复读取文件
+	a.cachedProjectContext = result
 	return result
 }
 
@@ -655,13 +662,17 @@ func (a *ConductorAgent) Run(ctx context.Context, input string, mem *memory.Conv
 		systemPrompt += "\nUse these agents via their delegate tools for tasks matching their specializations.\n"
 	}
 
-	// 加载项目上下文文件（CODEACTOR.md、CLAUDE.md、AGENTS.md）并前置到 System Prompt
-	if loadResult := a.loadProjectContext(); loadResult != nil && loadResult.Content != "" {
-		// 发送上下文加载完成消息到消息通道
-		if a.Publisher != nil {
-			a.Publisher.Publish("context_loaded", loadResult, a.Name())
+	// 只在首次对话时加载项目上下文文件（CODEACTOR.md、CLAUDE.md、AGENTS.md），
+	// 同一会话的后续追问无需重复注入，避免浪费 token。
+	// memory 中不存储 system 消息，因此 len(mem.GetMessages()) == 0 即可判断是否为首次对话。
+	if mem == nil || len(mem.GetMessages()) == 0 {
+		if loadResult := a.loadProjectContext(); loadResult != nil && loadResult.Content != "" {
+			// 发送上下文加载完成消息到消息通道
+			if a.Publisher != nil {
+				a.Publisher.Publish("context_loaded", loadResult, a.Name())
+			}
+			systemPrompt = fmt.Sprintf("### Project Workspace Context\n%s\n\n", loadResult.Content) + systemPrompt
 		}
-		systemPrompt = fmt.Sprintf("### Project Workspace Context\n%s\n\n", loadResult.Content) + systemPrompt
 	}
 
 	messages = append(messages, llm.Message{

internal/tui/render.go

@@ -101,7 +101,7 @@ func RenderToolLine(entry *ToolEntry, anim *Anim, width int) string {
 		params = formatToolParams(entry.Call.Name, entry.Call.Arguments)
 	}
 
-	// Check for early states first
+	// Check for early states first — no borders
 	if early, ok := RenderEarlyState(entry.Status, entry.Call.Name, params); ok {
 		return early
 	}
@@ -113,26 +113,35 @@ func RenderToolLine(entry *ToolEntry, anim *Anim, width int) string {
 
 	header := RenderHeader(entry.Status, entry.Call.Name, params, errBrief)
 
-	// If still running, this should have been handled by RenderPending — fallback
+	// If still running, no borders — animation is in progress
 	if entry.Status == ToolStatusRunning {
 		return header
 	}
 
 	// Tools whose result body is just status JSON — skip body rendering,
 	// only show the tool name + file path in the header.
 	if skipBodyTools[entry.Call.Name] && entry.Status == ToolStatusSuccess {
-		return header
+		return addToolCallBorders(header, width)
 	}
 
 	// Render body if we have a result
 	if entry.Result != nil && entry.Result.Content != "" {
 		body := RenderResultBody(entry.Result.Content, width)
 		if body != "" {
-			return header + "\n" + body
+			// Border only wraps header; body goes below the border
+			return addToolCallBorders(header, width) + "\n" + body
 		}
 	}
 
-	return header
+	return addToolCallBorders(header, width)
+}
+
+// addToolCallBorders wraps header with a thin top and bottom border line.
+// Body content should be appended OUTSIDE the border.
+func addToolCallBorders(header string, width int) string {
+	topBorder := ToolCallBorderTop.Render(makeBorderLine('─', width))
+	bottomBorder := ToolCallBorderBottom.Render(makeBorderLine('─', width))
+	return topBorder + "\n" + header + "\n" + bottomBorder
 }
 
 // RenderResultBody renders the tool result content with smart detection.
@@ -327,6 +336,15 @@ func renderCodeLines(content string, filename string, width int) string {
 
 // ── Content Detection ──
 
+// makeBorderLine generates a compact border line using half-block characters.
+func makeBorderLine(char rune, width int) string {
+	runes := make([]rune, width)
+	for i := range runes {
+		runes[i] = char
+	}
+	return string(runes)
+}
+
 func isJSON(s string) bool {
 	s = strings.TrimSpace(s)
 	if len(s) == 0 {

internal/tui/styles.go

@@ -107,6 +107,12 @@ var (
 	HelpStyle   = lipgloss.NewStyle().Foreground(lipgloss.Color("228"))
 )
 
+// ── Tool call area borders ──
+var (
+	ToolCallBorderTop    = lipgloss.NewStyle().Foreground(lipgloss.Color("240"))
+	ToolCallBorderBottom = lipgloss.NewStyle().Foreground(lipgloss.Color("240"))
+)
+
 // ── Separator ──
 var SeparatorStyle = lipgloss.NewStyle().Foreground(lipgloss.Color("237"))