Move hardcoded LLM prompts to template files (#3934)

Co-authored-by: Douwe Osinga <douwe@squareup.com>
Co-authored-by: Claude <noreply@anthropic.com>

crates/goose/src/permission/permission_judge.rs

@@ -3,6 +3,7 @@ use crate::config::permission::PermissionLevel;
 use crate::config::PermissionManager;
 use crate::conversation::message::{Message, MessageContent, ToolRequest};
 use crate::conversation::Conversation;
+use crate::prompt_template::render_global_file;
 use crate::providers::base::Provider;
 use chrono::Utc;
 use indoc::indoc;
@@ -13,6 +14,11 @@ use serde_json::Value;
 use std::collections::HashSet;
 use std::sync::Arc;
 
+#[derive(Serialize)]
+struct PermissionJudgeContext {
+    // Empty struct for now since the current template doesn't need variables
+}
+
 /// Creates the tool definition for checking read-only permissions.
 fn create_read_only_tool() -> Tool {
     Tool::new(
@@ -133,12 +139,12 @@ pub async fn detect_read_only_tools(
     let tool = create_read_only_tool();
     let check_messages = create_check_messages(tool_requests);
 
+    let context = PermissionJudgeContext {};
+    let system_prompt = render_global_file("permission_judge.md", &context)
+        .unwrap_or_else(|_| "You are a good analyst and can detect operations whether they have read-only operations.".to_string());
+
     let res = provider
-        .complete(
-            "You are a good analyst and can detect operations whether they have read-only operations.",
-            check_messages.messages(),
-            &[tool.clone()],
-        )
+        .complete(&system_prompt, check_messages.messages(), &[tool.clone()])
         .await;
 
     // Process the response and return an empty vector if the response is invalid