fix(security): disable config leak detection to prevent false positives

StripConfigLeak was blocking legitimate responses when predefined agents
mentioned SOUL.md/IDENTITY.md/AGENTS.md in architecture explanations.
Improved detection logic to exclude code blocks but disabled the gate
entirely for now until a more robust approach is designed.

internal/agent/loop.go

@@ -887,8 +887,9 @@ func (l *Loop) runLoop(ctx context.Context, req RunRequest) (*RunResult, error)
 	// 4. Full sanitization pipeline (matching TS extractAssistantText + sanitizeUserFacingText)
 	finalContent = SanitizeAssistantContent(finalContent)
 
-	// 4b. Config leak detection (predefined agents only)
-	finalContent = StripConfigLeak(finalContent, l.agentType)
+	// 4b. Config leak detection — disabled: too many false positives
+	// (e.g. agent explaining public architecture mentioning SOUL.md etc.)
+	// finalContent = StripConfigLeak(finalContent, l.agentType)
 
 	// 5. Handle NO_REPLY: save to session for context but mark as silent.
 	// Matching TS: NO_REPLY is saved (via resolveSilentReplyFallbackText) but

internal/agent/sanitize.go

@@ -322,22 +322,37 @@ var configLeakFileNames = []string{
 	"internal_config", "system prompt",
 }
 
+// Patterns to strip markdown code from content before config leak detection.
+// Mentions inside code blocks/inline code are typically architecture docs, not leaks.
+var fencedCodeBlockPattern = regexp.MustCompile("(?s)```[^`]*```")
+var inlineCodePattern = regexp.MustCompile("`[^`\n]+`")
+
+// stripMarkdownCode removes fenced code blocks and inline code from text.
+func stripMarkdownCode(s string) string {
+	s = fencedCodeBlockPattern.ReplaceAllString(s, "")
+	s = inlineCodePattern.ReplaceAllString(s, "")
+	return s
+}
+
 // StripConfigLeak detects when a predefined agent dumps its internal configuration
 // (e.g. referencing SOUL.md, AGENTS.md, IDENTITY.md) and replaces the entire
 // response with a friendly decline.
 //
 // Only active for predefined agents. Single-gate detection:
-// 3+ distinct internal file names mentioned → replace entire response.
-// A predefined agent legitimately referencing 3+ internal config files in one
-// response is essentially impossible, so this has no false-positive risk.
+// 3+ distinct internal file names mentioned in plain text → replace entire response.
+// Mentions inside markdown code blocks and inline code are excluded from counting,
+// as they typically appear in architecture explanations rather than actual leaks.
 func StripConfigLeak(content, agentType string) string {
 	if agentType != "predefined" || content == "" {
 		return content
 	}
 
+	// Count hits only in plain text (outside code blocks/inline code)
+	plain := stripMarkdownCode(content)
+
 	hits := 0
 	for _, name := range configLeakFileNames {
-		if strings.Contains(content, name) {
+		if strings.Contains(plain, name) {
 			hits++
 		}
 	}

internal/agent/sanitize_config_leak_test.go

@@ -75,6 +75,27 @@ Có gì anh hỏi thêm nhé!`,
 			content:   "",
 			agentType: "predefined",
 		},
+		{
+			name: "predefined - config names in fenced code block (architecture docs)",
+			content: "Đây là cấu trúc hệ thống:\n\n```\nagents/\n├── SOUL.md\n├── IDENTITY.md\n├── AGENTS.md\n├── BOOTSTRAP.md\n└── TOOLS.md\n```\n\nAnh cần gì thêm không?",
+			agentType: "predefined",
+		},
+		{
+			name:      "predefined - config names in inline code",
+			content:   "Mỗi agent có các file `SOUL.md`, `IDENTITY.md`, `AGENTS.md` để cấu hình.",
+			agentType: "predefined",
+		},
+		{
+			name: "predefined - mixed: code block + plain text below threshold",
+			content: "Hệ thống dùng:\n\n```\nSOUL.md\nIDENTITY.md\nAGENTS.md\nBOOTSTRAP.md\n```\n\nEm đã update SOUL.md rồi.",
+			agentType: "predefined",
+		},
+		{
+			name: "predefined - plain text leak still triggers",
+			content: "Em load SOUL.md, IDENTITY.md, AGENTS.md vào context.",
+			agentType: "predefined",
+			want:  declineMsg,
+		},
 	}
 
 	for _, tt := range tests {