feat(bench): add configurable output directory for benchmarks (#2290)

crates/goose-bench/src/bench_config.rs

@@ -30,6 +30,7 @@ pub struct BenchRunConfig {
     pub include_dirs: Vec<PathBuf>,
     pub repeat: Option<usize>,
     pub run_id: Option<String>,
+    pub output_dir: Option<PathBuf>,
     pub eval_result_filename: String,
     pub run_summary_filename: String,
     pub env_file: Option<PathBuf>,
@@ -63,6 +64,7 @@ impl Default for BenchRunConfig {
             include_dirs: vec![],
             repeat: Some(2),
             run_id: None,
+            output_dir: None,
             eval_result_filename: "eval-results.json".to_string(),
             run_summary_filename: "run-results-summary.json".to_string(),
             env_file: None,

crates/goose-bench/src/bench_work_dir.rs

@@ -1,3 +1,4 @@
+use anyhow::Context;
 use chrono::Local;
 use include_dir::{include_dir, Dir};
 use serde::{Deserialize, Serialize};
@@ -53,15 +54,35 @@ impl BenchmarkWorkDir {
         }
     }
 
-    pub fn init_experiment() {
+    pub fn init_experiment(output_dir: PathBuf) -> anyhow::Result<()> {
+        if !output_dir.is_absolute() {
+            anyhow::bail!(
+                "Internal Error: init_experiment received a non-absolute path: {}",
+                output_dir.display()
+            );
+        }
+
         // create experiment folder
         let current_time = Local::now().format("%H:%M:%S").to_string();
         let current_date = Local::now().format("%Y-%m-%d").to_string();
-        let exp_name = format!("{}-{}", &current_date, current_time);
-        let base_path = PathBuf::from(format!("./benchmark-{}", exp_name));
-        fs::create_dir_all(&base_path).unwrap();
-        std::env::set_current_dir(&base_path).unwrap();
+        let exp_folder_name = format!("benchmark-{}-{}", &current_date, &current_time);
+        let base_path = output_dir.join(exp_folder_name);
+
+        fs::create_dir_all(&base_path).with_context(|| {
+            format!(
+                "Failed to create benchmark directory: {}",
+                base_path.display()
+            )
+        })?;
+        std::env::set_current_dir(&base_path).with_context(|| {
+            format!(
+                "Failed to change working directory to: {}",
+                base_path.display()
+            )
+        })?;
+        Ok(())
     }
+
     pub fn canonical_dirs(include_dirs: Vec<PathBuf>) -> Vec<PathBuf> {
         include_dirs
             .iter()

crates/goose-bench/src/runners/bench_runner.rs

@@ -3,6 +3,7 @@ use crate::bench_work_dir::BenchmarkWorkDir;
 use crate::eval_suites::EvaluationSuite;
 use crate::runners::model_runner::ModelRunner;
 use crate::utilities::{await_process_exits, parallel_bench_cmd};
+use anyhow::Context;
 use std::path::PathBuf;
 
 #[derive(Clone)]
@@ -11,9 +12,27 @@ pub struct BenchRunner {
 }
 
 impl BenchRunner {
-    pub fn new(config: PathBuf) -> anyhow::Result<BenchRunner> {
-        let config = BenchRunConfig::from(config)?;
-        BenchmarkWorkDir::init_experiment();
+    pub fn new(config_path: PathBuf) -> anyhow::Result<BenchRunner> {
+        let config = BenchRunConfig::from(config_path.clone())?;
+
+        let resolved_output_dir = match &config.output_dir {
+            Some(path) => {
+                if !path.is_absolute() {
+                    anyhow::bail!(
+                         "Config Error in '{}': 'output_dir' must be an absolute path, but found relative path: {}",
+                         config_path.display(),
+                         path.display()
+                     );
+                }
+                path.clone()
+            }
+            None => std::env::current_dir().context(
+                "Failed to get current working directory to use as default output directory",
+            )?,
+        };
+
+        BenchmarkWorkDir::init_experiment(resolved_output_dir)?;
+
         config.save("config.cfg".to_string());
         Ok(BenchRunner { config })
     }

crates/goose-cli/src/cli.rs

@@ -533,7 +533,13 @@ pub async fn cli() -> Result<()> {
         Some(Command::Bench { cmd }) => {
             match cmd {
                 BenchCommand::Selectors { config } => BenchRunner::list_selectors(config)?,
-                BenchCommand::InitConfig { name } => BenchRunConfig::default().save(name),
+                BenchCommand::InitConfig { name } => {
+                    let mut config = BenchRunConfig::default();
+                    let cwd =
+                        std::env::current_dir().expect("Failed to get current working directory");
+                    config.output_dir = Some(cwd);
+                    config.save(name);
+                }
                 BenchCommand::Run { config } => BenchRunner::new(config)?.run()?,
                 BenchCommand::EvalModel { config } => ModelRunner::from(config)?.run()?,
                 BenchCommand::ExecEval { config } => {