tidy: clean up old benchmark and add gym (#7081)

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

@@ -1,76 +0,0 @@
-use crate::cli::StreamableHttpOptions;
-use crate::session::build_session;
-use crate::session::SessionBuilderConfig;
-use crate::{logging, CliSession};
-use async_trait::async_trait;
-use goose::conversation::Conversation;
-use goose::session::session_manager::Session;
-use goose_bench::bench_session::{BenchAgent, BenchBaseSession};
-use goose_bench::eval_suites::ExtensionRequirements;
-use std::sync::Arc;
-use tokio::sync::Mutex;
-
-// allow session obj to be used in benchmarking
-#[async_trait]
-impl BenchBaseSession for CliSession {
-    async fn headless(&mut self, message: String) -> anyhow::Result<()> {
-        self.headless(message).await
-    }
-    fn message_history(&self) -> Conversation {
-        self.message_history()
-    }
-    fn get_total_token_usage(&self) -> anyhow::Result<Option<i32>> {
-        // Since the trait requires sync but the session method is async,
-        // we need to block on the async call
-        tokio::task::block_in_place(|| {
-            tokio::runtime::Handle::current().block_on(self.get_total_token_usage())
-        })
-    }
-
-    async fn get_session(&self) -> anyhow::Result<Session> {
-        self.get_session().await
-    }
-}
-pub async fn agent_generator(
-    requirements: ExtensionRequirements,
-    session_id: String,
-) -> BenchAgent {
-    let streamable_http_extensions: Vec<StreamableHttpOptions> = requirements
-        .streamable_http
-        .iter()
-        .map(|s| StreamableHttpOptions {
-            url: s.clone(),
-            timeout: goose::config::DEFAULT_EXTENSION_TIMEOUT,
-        })
-        .collect();
-    let base_session = build_session(SessionBuilderConfig {
-        session_id: Some(session_id),
-        resume: false,
-        fork: false,
-        no_session: false,
-        extensions: requirements.external,
-        streamable_http_extensions,
-        builtins: requirements.builtin,
-        no_profile: true,
-        recipe: None,
-        additional_system_prompt: None,
-        provider: None,
-        model: None,
-        debug: false,
-        max_tool_repetitions: None,
-        interactive: false, // Benchmarking is non-interactive
-        scheduled_job_id: None,
-        max_turns: None,
-        quiet: false,
-        output_format: "text".to_string(),
-        container: None,
-    })
-    .await;
-
-    let bench_agent = BenchAgent::new(Box::new(base_session));
-
-    let errors = Some(Arc::new(Mutex::new(bench_agent.get_errors().await)));
-    logging::setup_logging(Some("bench"), errors).expect("Failed to initialize logging");
-
-    bench_agent
-}

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

@@ -1,4 +1,3 @@
-pub mod bench;
 pub mod configure;
 pub mod info;
 pub mod project;

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

@@ -238,7 +238,7 @@ pub async fn handle_web(
     no_auth: bool,
 ) -> Result<()> {
     validate_network_auth(&host, &auth_token, no_auth);
-    crate::logging::setup_logging(Some("goose-web"), None)?;
+    crate::logging::setup_logging(Some("goose-web"))?;
 
     let (provider_name, model) = get_provider_and_model();
     let agent = create_agent(&provider_name, &model).await?;