From d76f8b4c418c164ca8e50fdc1748fe98a06d8488 Mon Sep 17 00:00:00 2001
From: Marcelle Bonterre <marcelle@squareup.com>
Date: Thu, 20 Feb 2025 09:40:05 -0500
Subject: [PATCH] bugfix the flattening

---
 crates/goose-cli/src/commands/bench.rs | 38 ++++++++++++++++----------
 crates/goose-cli/src/main.rs           |  2 +-
 2 files changed, 25 insertions(+), 15 deletions(-)

diff --git a/crates/goose-cli/src/commands/bench.rs b/crates/goose-cli/src/commands/bench.rs
index 1b5df58ae..b7826b866 100644
--- a/crates/goose-cli/src/commands/bench.rs
+++ b/crates/goose-cli/src/commands/bench.rs
@@ -4,7 +4,7 @@ use async_trait::async_trait;
 use chrono::Local;
 use goose::config::Config;
 use goose::message::Message;
-use goose_bench::eval_suites::{BenchAgent, EvaluationMetric, EvaluationSuiteFactory};
+use goose_bench::eval_suites::{BenchAgent, Evaluation, EvaluationMetric, EvaluationSuiteFactory};
 use goose_bench::work_dir::WorkDir;
 
 #[async_trait]
@@ -15,23 +15,31 @@ impl BenchAgent for Session {
     }
 }
 
-async fn run_eval(mut evaluation: Box<dyn BenchAgent>) -> anyhow::Result<Vec<EvaluationMetric>> {
-    let _ = WorkDir::work_from(format!("./{}", &evaluation.name()));
-    let session = build_session(None, false, Vec::new(), Vec::new()).await;
-    let report = evaluation.run(Box::new(session))?.await;
-    report
+async fn run_eval(evaluation: Box<dyn Evaluation>) -> anyhow::Result<Vec<EvaluationMetric>> {
+    if let Ok(_) = WorkDir::work_from(format!("./{}", &evaluation.name())) {
+        let session = build_session(None, false, Vec::new(), Vec::new()).await;
+        let report = evaluation.run(Box::new(session)).await;
+        report
+    }else{
+        Ok(vec![])
+    }
 }
 
 async fn run_suite(suite: &str, current_time: &String, current_date: &String) -> anyhow::Result<()> {
-    let _ = WorkDir::work_from(format!("./{}", &suite))?;
-    let _ = WorkDir::work_from(format!("./{}-{}", &current_date, current_time))?;
-    for Some(evaluation) in EvaluationSuiteFactory::create(suite) {
-        run_eval(evaluation)?.await;
+    if let Ok(_) = WorkDir::work_from(format!("./{}", &suite)) {
+        if let Ok(_) = WorkDir::work_from(format!("./{}-{}", &current_date, current_time)) {
+            if let Some(evals) = EvaluationSuiteFactory::create(suite) {
+                for eval in evals {
+                    run_eval(eval).await?;
+                }
+            }
+        }
     }
+
     Ok(())
 }
 
-pub async fn run_benchmark(suites: Vec<String>) {
+pub async fn run_benchmark(suites: Vec<String>) -> anyhow::Result<()> {
     let suites = EvaluationSuiteFactory::available_evaluations()
         .into_iter()
         .filter(|&s| suites.contains(&s.to_string()))
@@ -45,8 +53,10 @@ pub async fn run_benchmark(suites: Vec<String>) {
 
     let current_time = Local::now().format("%H:%M:%S").to_string();
     let current_date = Local::now().format("%Y-%m-%d").to_string();
-    let _ = WorkDir::work_from(format!("./benchmark-{}", &provider_name))?;
-    for suite in suites {
-        run_suite(suite, &current_time, &current_date)?.await;
+    if let Ok(_) = WorkDir::work_from(format!("./benchmark-{}", &provider_name)) {
+        for suite in suites {
+            run_suite(suite, &current_time, &current_date).await?;
+        }
     }
+    Ok(())
 }
diff --git a/crates/goose-cli/src/main.rs b/crates/goose-cli/src/main.rs
index e53883b83..becb1d93d 100644
--- a/crates/goose-cli/src/main.rs
+++ b/crates/goose-cli/src/main.rs
@@ -227,7 +227,7 @@ async fn main() -> Result<()> {
             } else {
                 suites
             };
-            run_benchmark(suites).await;
+            let _ = run_benchmark(suites).await;
             return Ok(());
         }
         None => {