feat: 加固 Agent CLI 执行边界并更新模型目录

收紧 SSH 命令护栏、权限校验与远程执行会话，并同步主流大模型供应商列表。

Co-authored-by: Cursor <cursoragent@cursor.com>

backend/pkg/agentcli/core/agent.go

@@ -13,6 +13,7 @@ import (
 	"time"
 
 	"ezops/pkg/agentcli/evolver"
+	"ezops/pkg/agentcli/guardrail"
 	"ezops/pkg/agentcli/llm"
 	"ezops/pkg/agentcli/prompt"
 	"ezops/pkg/agentcli/safety"
@@ -293,9 +294,10 @@ func (a *Agent) processToolCalls(ctx context.Context, calls []llm.ToolCall, hist
 				Content: "错误: 参数解析失败: " + err.Error()})
 			continue
 		}
+		safeArgs := guardrail.SanitizeMap(args)
 
 		if msg := a.suppressBroadKubernetesGet(tc.Function.Name, args, history); msg != "" {
-			a.emit(Event{Type: EventToolCall, Phase: PhaseToolRunning, Round: round, ToolName: tc.Function.Name, ToolArgs: args, Risk: tools.RiskSafe.Display(), Summary: "工具执行中"})
+			a.emit(Event{Type: EventToolCall, Phase: PhaseToolRunning, Round: round, ToolName: tc.Function.Name, ToolArgs: safeArgs, Risk: tools.RiskSafe.Display(), Summary: "工具执行中"})
 			a.emit(Event{Type: EventToolResult, Phase: PhaseObserving, Round: round, ToolName: tc.Function.Name,
 				Output: msg, Success: true, Duration: "0s", OutputSize: len(msg)})
 			obs = append(obs, llm.Message{Role: "tool", ToolCallID: tc.ID, Name: tc.Function.Name, Content: msg})
@@ -303,12 +305,12 @@ func (a *Agent) processToolCalls(ctx context.Context, calls []llm.ToolCall, hist
 		}
 
 		risk := a.toolRisk(tool, args)
-		a.emit(Event{Type: EventStatus, Phase: PhaseSafety, Round: round, ToolName: tc.Function.Name, ToolArgs: args, Risk: risk.Display(), Summary: "安全检查"})
+		a.emit(Event{Type: EventStatus, Phase: PhaseSafety, Round: round, ToolName: tc.Function.Name, ToolArgs: safeArgs, Risk: risk.Display(), Summary: "安全检查"})
 
 		// 安全检查（高危触发 confirmFn，由 controller 异步确认）
 		approved, err := a.safetyCtl.Check(tool, args)
 		if err != nil {
-			a.emit(Event{Type: EventStatus, Phase: PhaseError, Round: round, ToolName: tc.Function.Name, ToolArgs: args, Risk: risk.Display(), Message: err.Error()})
+			a.emit(Event{Type: EventStatus, Phase: PhaseError, Round: round, ToolName: tc.Function.Name, ToolArgs: safeArgs, Risk: risk.Display(), Message: guardrail.SanitizeText(err.Error())})
 			obs = append(obs, llm.Message{Role: "tool", ToolCallID: tc.ID, Name: tc.Function.Name,
 				Content: "安全检查失败: " + err.Error()})
 			continue
@@ -316,19 +318,19 @@ func (a *Agent) processToolCalls(ctx context.Context, calls []llm.ToolCall, hist
 		if !approved {
 			obs = append(obs, llm.Message{Role: "tool", ToolCallID: tc.ID, Name: tc.Function.Name,
 				Content: "操作已被用户拒绝执行"})
-			a.emit(Event{Type: EventStatus, Phase: PhaseBlocked, Round: round, ToolName: tc.Function.Name, ToolArgs: args, Risk: risk.Display(), Summary: "用户拒绝执行"})
+			a.emit(Event{Type: EventStatus, Phase: PhaseBlocked, Round: round, ToolName: tc.Function.Name, ToolArgs: safeArgs, Risk: risk.Display(), Summary: "用户拒绝执行"})
 			a.emit(Event{Type: EventToolResult, Round: round, ToolName: tc.Function.Name,
 				Phase: PhaseObserving, Output: "（用户拒绝执行）", Success: false, OutputSize: len("（用户拒绝执行）")})
 			continue
 		}
 
 		// 执行
-		a.emit(Event{Type: EventToolCall, Phase: PhaseToolRunning, Round: round, ToolName: tc.Function.Name, ToolArgs: args, Risk: risk.Display(), Summary: "工具执行中"})
+		a.emit(Event{Type: EventToolCall, Phase: PhaseToolRunning, Round: round, ToolName: tc.Function.Name, ToolArgs: safeArgs, Risk: risk.Display(), Summary: "工具执行中"})
 		if cached, ok := a.cachedToolResult(tc.Function.Name, args); ok {
 			a.emit(Event{Type: EventToolResult, Phase: PhaseObserving, Round: round, ToolName: tc.Function.Name,
 				Output: cached, Success: true, Duration: "0s", OutputSize: len(cached)})
 			records = append(records, evolver.ToolCallRecord{
-				ToolName: tc.Function.Name, Args: args, Output: cached,
+				ToolName: tc.Function.Name, Args: safeArgs, Output: cached,
 				Duration: 0, Success: true, RiskLevel: int(risk),
 			})
 			obs = append(obs, llm.Message{Role: "tool", ToolCallID: tc.ID, Name: tc.Function.Name, Content: cached})
@@ -349,6 +351,9 @@ func (a *Agent) processToolCalls(ctx context.Context, calls []llm.ToolCall, hist
 			errMsg = execErr.Error()
 			success = false
 		}
+		// 工具结果必须先脱敏，再进入自动分析、事件、缓存、Evolver 和模型上下文。 // Logan.Li
+		output = guardrail.SanitizeText(output)
+		errMsg = guardrail.SanitizeText(errMsg)
 
 		// 超长命令输出自动分析，避免 token 爆炸 // Logan.Li
 		if (tc.Function.Name == "ssh_execute" || tc.Function.Name == "pod_execute") && success && len(output) > a.autoAnalyzeThreshold() {
@@ -365,10 +370,10 @@ func (a *Agent) processToolCalls(ctx context.Context, calls []llm.ToolCall, hist
 
 		a.emit(Event{Type: EventToolResult, Phase: PhaseObserving, Round: round, ToolName: tc.Function.Name,
 			Output: output, Success: success, Duration: dur.Truncate(time.Millisecond).String(), OutputSize: len(output)})
-		a.storeToolResult(tc.Function.Name, args, output, success)
+		a.storeToolResult(tc.Function.Name, safeArgs, output, success)
 
 		records = append(records, evolver.ToolCallRecord{
-			ToolName: tc.Function.Name, Args: args, Output: output,
+			ToolName: tc.Function.Name, Args: safeArgs, Output: output,
 			Duration: dur, Success: success, RiskLevel: int(risk),
 		})
 
@@ -376,7 +381,7 @@ func (a *Agent) processToolCalls(ctx context.Context, calls []llm.ToolCall, hist
 		var content string
 		switch {
 		case execErr != nil:
-			content = "执行错误: " + execErr.Error()
+			content = "执行错误: " + errMsg
 			failInc++
 		case !success:
 			content = fmt.Sprintf("执行失败: %s\n输出: %s", errMsg, output)

backend/pkg/agentcli/guardrail/sanitize.go

@@ -15,7 +15,7 @@ var sensitivePatterns = []*regexp.Regexp{
 	regexp.MustCompile(`(?i)((?:mysql|postgres|postgresql|redis|mongodb)://[^:\s/@]+:)[^@\s]+(@)`),
 	regexp.MustCompile(`(?is)-----BEGIN [A-Z ]*PRIVATE KEY-----.*?-----END [A-Z ]*PRIVATE KEY-----`),
 	regexp.MustCompile(`(?i)((?:Cookie|Set-Cookie):\s*)[^\r\n]+`),
-	regexp.MustCompile(`(?i)(AKIA[0-9A-Z]{16})`),
+	regexp.MustCompile(`(?i)AKIA[0-9A-Z]{16}`),
 }
 
 var sensitiveKeys = map[string]bool{
@@ -65,6 +65,18 @@ func SanitizeJSON(raw string) string {
 	return string(b)
 }
 
+// SanitizeMap 返回脱敏后的参数副本，避免事件、审计和记忆泄露工具参数中的凭证。 // Logan.Li
+func SanitizeMap(input map[string]interface{}) map[string]interface{} {
+	if input == nil {
+		return nil
+	}
+	sanitized, ok := sanitizeValue(input).(map[string]interface{})
+	if !ok {
+		return map[string]interface{}{}
+	}
+	return sanitized
+}
+
 func sanitizeValue(v interface{}) interface{} {
 	switch val := v.(type) {
 	case map[string]interface{}:

backend/pkg/agentcli/guardrail/sanitize_test.go

@@ -6,9 +6,9 @@ import (
 )
 
 func TestSanitizeTextMasksCommonSecrets(t *testing.T) {
-	input := "Authorization: Bearer abcdefghijklmnopqrstuvwxyz password=secret123 mysql://root:pass@db/prod"
+	input := "Authorization: Bearer abcdefghijklmnopqrstuvwxyz password=secret123 mysql://root:pass@db/prod AKIA1234567890ABCDEF"
 	got := SanitizeText(input)
-	for _, leak := range []string{"abcdefghijklmnopqrstuvwxyz", "secret123", "root:pass@db"} {
+	for _, leak := range []string{"abcdefghijklmnopqrstuvwxyz", "secret123", "root:pass@db", "AKIA1234567890ABCDEF"} {
 		if strings.Contains(got, leak) {
 			t.Fatalf("敏感信息未脱敏: %s in %s", leak, got)
 		}
@@ -29,3 +29,22 @@ func TestSanitizeJSONMasksSensitiveKeys(t *testing.T) {
 		t.Fatalf("非敏感字段不应被改写: %s", got)
 	}
 }
+
+func TestSanitizeMapReturnsMaskedCopy(t *testing.T) {
+	input := map[string]interface{}{
+		"command": "curl -H 'Authorization: Bearer abcdefghijklmnop' https://example.test",
+		"nested":  map[string]interface{}{"password": "plain-secret"},
+	}
+
+	got := SanitizeMap(input)
+	if got["command"] == input["command"] {
+		t.Fatal("命令中的令牌未脱敏")
+	}
+	nested := got["nested"].(map[string]interface{})
+	if nested["password"] != "[REDACTED]" {
+		t.Fatalf("密码字段未脱敏: %#v", nested["password"])
+	}
+	if input["nested"].(map[string]interface{})["password"] != "plain-secret" {
+		t.Fatal("脱敏不得修改原始参数")
+	}
+}

backend/pkg/agentcli/tools/ssh_execute.go

@@ -19,6 +19,8 @@ type SSHExecuteTool struct {
 	timeout time.Duration                // 默认命令超时
 }
 
+const maxSSHCommandTimeout = 300 * time.Second
+
 // NewSSHExecuteTool 创建远程执行工具（内部建立 SSH 长连接） // Logan.Li
 func NewSSHExecuteTool(cfg remoteexec.SSHConfig, timeout time.Duration) (*SSHExecuteTool, error) {
 	if timeout <= 0 {
@@ -82,12 +84,12 @@ func (t *SSHExecuteTool) Execute(ctx context.Context, args map[string]interface{
 		return &Result{Success: false, Error: "参数 command 不能为空"}, nil
 	}
 
-	timeout := t.timeout
+	timeout := boundedSSHTimeout(t.timeout)
 	if sec := IntParam(args, "timeout"); sec > 0 {
-		timeout = time.Duration(sec) * time.Second
+		timeout = boundedSSHTimeout(time.Duration(sec) * time.Second)
 	}
 
-	res, err := t.client.Exec(command, timeout)
+	res, err := t.client.ExecContext(ctx, command, timeout)
 	if err != nil {
 		return &Result{Success: false, Error: fmt.Sprintf("SSH 连接失败: %v", err)}, nil
 	}
@@ -103,4 +105,15 @@ func (t *SSHExecuteTool) Execute(ctx context.Context, args map[string]interface{
 		Output:  output,
 		Error:   res.Error,
 	}, nil
-}
+}
+
+// boundedSSHTimeout 将模型提供的超时限制在平台允许范围内。 // Logan.Li
+func boundedSSHTimeout(timeout time.Duration) time.Duration {
+	if timeout <= 0 {
+		return time.Second
+	}
+	if timeout > maxSSHCommandTimeout {
+		return maxSSHCommandTimeout
+	}
+	return timeout
+}

backend/pkg/agentcli/tools/ssh_execute_test.go

@@ -0,0 +1,25 @@
+package tools
+
+import (
+	"testing"
+	"time"
+)
+
+func TestBoundedSSHTimeout(t *testing.T) {
+	cases := []struct {
+		name string
+		in   time.Duration
+		want time.Duration
+	}{
+		{name: "non-positive", in: 0, want: time.Second},
+		{name: "normal", in: 60 * time.Second, want: 60 * time.Second},
+		{name: "too-large", in: 24 * time.Hour, want: 300 * time.Second},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := boundedSSHTimeout(tc.in); got != tc.want {
+				t.Fatalf("boundedSSHTimeout(%v)=%v, want %v", tc.in, got, tc.want)
+			}
+		})
+	}
+}