Feat: support auto-including dirs in binary/bench-work-dir (#1576)

Cargo.lock

@@ -2254,6 +2254,7 @@ dependencies = [
  "chrono",
  "ctor",
  "goose",
+ "include_dir",
  "mcp-core",
  "paste",
  "serde",

crates/goose-bench/Cargo.toml

@@ -21,6 +21,7 @@ serde = { version = "1.0", features = ["derive"] }
 tracing = "0.1"
 tracing-subscriber = { version = "0.3", features = ["registry"] }
 tokio = { version = "1.0", features = ["full"] }
+include_dir = "0.7.4"
 
 [target.'cfg(target_os = "windows")'.dependencies]
 winapi = { version = "0.3", features = ["wincred"] }

crates/goose-bench/src/bench_work_dir.rs

@@ -1,4 +1,5 @@
 use chrono::Local;
+use include_dir::{include_dir, Dir};
 use std::fs;
 use std::io;
 use std::io::ErrorKind;
@@ -6,6 +7,8 @@ use std::path::Path;
 use std::path::PathBuf;
 use std::process::Command;
 
+pub static BUILTIN_EVAL_ASSETS: Dir = include_dir!("$CARGO_MANIFEST_DIR/src");
+
 pub struct BenchmarkWorkDir {
     pub base_path: PathBuf,
     cwd: PathBuf,
@@ -16,7 +19,7 @@ pub struct BenchmarkWorkDir {
 
 impl Default for BenchmarkWorkDir {
     fn default() -> Self {
-        BenchmarkWorkDir::new("work_dir".to_string(), Vec::new())
+        Self::new("work_dir".to_string(), Vec::new())
     }
 }
 impl BenchmarkWorkDir {
@@ -42,9 +45,11 @@ impl BenchmarkWorkDir {
         // deep copy each dir
         let _: Vec<_> = dirs
             .iter()
-            .map(|d| BenchmarkWorkDir::deep_copy(d.as_path(), base_path.as_path(), true))
+            .map(|d| Self::deep_copy(d.as_path(), base_path.as_path(), true))
             .collect();
 
+        Self::copy_auto_included_dirs(&base_path);
+
         std::env::set_current_dir(&base_path).unwrap();
 
         BenchmarkWorkDir {
@@ -55,6 +60,13 @@ impl BenchmarkWorkDir {
             eval: None,
         }
     }
+    fn copy_auto_included_dirs(dest: &PathBuf) {
+        BUILTIN_EVAL_ASSETS
+            .get_dir("assets")
+            .unwrap()
+            .extract(dest)
+            .unwrap();
+    }
     pub fn cd(&mut self, path: PathBuf) -> anyhow::Result<&mut Self> {
         fs::create_dir_all(&path)?;
         std::env::set_current_dir(&path)?;
@@ -132,7 +144,7 @@ impl BenchmarkWorkDir {
         let here = PathBuf::from(".").canonicalize()?;
         let artifact_at_root = self.base_path.clone().join(asset_rel_path);
 
-        BenchmarkWorkDir::deep_copy(artifact_at_root.as_path(), here.as_path(), true)?;
+        Self::deep_copy(artifact_at_root.as_path(), here.as_path(), true)?;
         Ok(PathBuf::from(path))
     }
 

crates/goose-cli/src/commands/bench.rs

@@ -115,7 +115,7 @@ async fn run_eval(
 
 async fn run_suite(suite: &str, work_dir: &mut BenchmarkWorkDir) -> anyhow::Result<SuiteResult> {
     let mut suite_result = SuiteResult::new(suite.to_string());
-    let eval_lock = Mutex::new(0);
+    let eval_lock = Mutex::new(());
 
     if let Some(evals) = EvaluationSuiteFactory::create(suite) {
         for eval in evals {
@@ -152,7 +152,7 @@ pub async fn run_benchmark(
         format!("{}-{}", provider_name, goose_model),
         include_dirs.clone(),
     );
-    let suite_lock = Mutex::new(0);
+    let suite_lock = Mutex::new(());
     for suite in suites {
         let _unused = suite_lock.lock().await;
         work_dir.set_suite(suite);