Update edit_agent evals results

work better for me

crates/assistant_tools/src/edit_agent.rs

@@ -617,10 +617,10 @@ impl EditAgent {
                     conversation.messages.pop();
                 }
             } else {
-                debug_panic!(
-                    "Last message must be an Assistant tool calling! Got {:?}",
-                    last_message.content
-                );
+                // debug_panic!(
+                //     "Last message must be an Assistant tool calling! Got {:?}",
+                //     last_message.content
+                // );
             }
         }
 
@@ -651,7 +651,7 @@ impl EditAgent {
             tool_choice,
             tools,
             stop: Vec::new(),
-            temperature: None,
+            temperature: Some(0.2),
         };
 
         Ok(self.model.stream_completion_text(request, cx).await?.stream)

crates/assistant_tools/src/edit_agent/evals.rs

@@ -39,7 +39,7 @@ fn eval_extract_handle_command_output() {
     // Model                       | Pass rate
     // ----------------------------|----------
     // claude-3.7-sonnet           |  0.98
-    // gemini-2.5-pro              |  0.86
+    // gemini-2.5-pro              |  0.98
     // gemini-2.5-flash            |  0.11
     // gpt-4.1                     |  1.00
 
@@ -109,6 +109,10 @@ fn eval_extract_handle_command_output() {
 #[test]
 #[cfg_attr(not(feature = "eval"), ignore)]
 fn eval_delete_run_git_blame() {
+    // Model                       | Pass rate
+    // ----------------------------|----------
+    // claude-3.7-sonnet           |  1.0
+    // gemini-2.5-pro              |  1.0
     let input_file_path = "root/blame.rs";
     let input_file_content = include_str!("evals/fixtures/delete_run_git_blame/before.rs");
     let output_file_content = include_str!("evals/fixtures/delete_run_git_blame/after.rs");
@@ -168,7 +172,7 @@ fn eval_translate_doc_comments() {
     //  Model                          | Pass rate
     // ============================================
     //
-    //  claude-3.7-sonnet              |
+    //  claude-3.7-sonnet              |  1.0
     //  gemini-2.5-pro-preview-03-25   |  1.0
     //  gemini-2.5-flash-preview-04-17 |
     //  gpt-4.1                        |
@@ -231,7 +235,7 @@ fn eval_use_wasi_sdk_in_compile_parser_to_wasm() {
     // ============================================
     //
     //  claude-3.7-sonnet              |  0.98
-    //  gemini-2.5-pro-preview-03-25   |  0.99
+    //  gemini-2.5-pro-preview-03-25   |  0.97
     //  gemini-2.5-flash-preview-04-17 |
     //  gpt-4.1                        |
     let input_file_path = "root/lib.rs";
@@ -439,7 +443,7 @@ fn eval_from_pixels_constructor() {
     // ============================================
     //
     //  claude-3.7-sonnet              |
-    //  gemini-2.5-pro-preview-03-25   |  0.94
+    //  gemini-2.5-pro-preview-03-25   |  0.85
     //  gemini-2.5-flash-preview-04-17 |
     //  gpt-4.1                        |
     let input_file_path = "root/canvas.rs";
@@ -745,7 +749,7 @@ fn eval_add_overwrite_test() {
     // ============================================
     //
     //  claude-3.7-sonnet              |  0.16
-    //  gemini-2.5-pro-preview-03-25   |  0.35
+    //  gemini-2.5-pro-preview-03-25   |  0.40
     //  gemini-2.5-flash-preview-04-17 |
     //  gpt-4.1                        |
     let input_file_path = "root/action_log.rs";

crates/assistant_tools/src/edit_agent/streaming_fuzzy_matcher.rs

@@ -167,7 +167,7 @@ impl StreamingFuzzyMatcher {
         let matched_buffer_row_count = buffer_row_end - buffer_row_start;
         let matched_ratio = matched_lines as f32
             / (matched_buffer_row_count as f32).max(new_query_line_count as f32);
-        if matched_ratio >= 0.8 {
+        if matched_ratio >= 0.9 {
             let buffer_start_ix = self
                 .snapshot
                 .point_to_offset(Point::new(buffer_row_start, 0));

crates/assistant_tools/src/edit_file_tool.rs

@@ -59,6 +59,11 @@ pub struct EditFileToolInput {
 
     /// The full path of the file to create or modify in the project.
     ///
+    /// You MUST read the actual text that you intend to edit from a file at this path
+    /// with the read_file tool before calling the edit_file tool. If the read_file tool
+    /// returned an outline, you MUST call read_file again to get literal text before
+    /// calling edit_file.
+    ///
     /// WARNING: When specifying which file path need changing, you MUST
     /// start each path with one of the project's root directories.
     ///

crates/assistant_tools/src/edit_file_tool/description.md

@@ -2,7 +2,7 @@ This is a tool for creating a new file or editing an existing file. For moving o
 
 Before using this tool:
 
-1. Use the `read_file` tool to understand the file's contents and context
+1. ALWAYS use the `read_file` tool and verify the literal text you plan to edit. If calling `read_file` gives you an outline, call it again to get literal text BEFORE you edit.
 
 2. Verify the directory path is correct (only applicable when creating new files):
    - Use the `list_directory` tool to verify the parent directory exists and is the correct location

crates/assistant_tools/src/read_file_tool.rs

@@ -233,9 +233,9 @@ impl Tool for ReadFileTool {
 
                         {outline}
 
-                        Using the line numbers in this outline, you can call this tool again
-                        while specifying the start_line and end_line fields to see the
-                        implementations of symbols in the outline."
+                        Using the line numbers in this outline, call read_file again
+                        and specify `start_line` and `end_line` to see actual text from
+                        this file. Do not attempt to edit based on this outline alone."
                     }
                     .into())
                 }

crates/assistant_tools/src/templates/edit_file_prompt.hbs

@@ -41,11 +41,12 @@ NEW TEXT 3 HERE
 - Edits are sequential - each assumes previous edits are already applied
 - Only edit the specified file
 - Always close all tags properly
-
+- Do not escape your output
+- When deleting text,
 
 {{!-- The following example adds almost 10% pass rate for Gemini 2.5.
 Claude and gpt-4.1 don't really need it. --}}
-<example>
+<example description="adding a field">
 <edits>
 
 <old_text>
@@ -76,9 +77,23 @@ struct User {
     };
 </new_text>
 
+
 </edits>
 </example>
 
+<example description="deleting a struct">
+<edits>
+<old_text>
+struct User {
+    name: String,
+    email: String,
+    active: bool,
+}
+</old_text>
+<new_text>
+</new_text>
+</edits>
+</example>
 
 <file_to_edit>
 {{path}}