nit: show dir in title, and less... jank (#7138)

crates/goose-cli/src/session/completion.rs

@@ -410,10 +410,7 @@ impl Hinter for GooseCompleter {
             }
             HintStatus::Default => {
                 let newline_key = super::input::get_newline_key().to_ascii_uppercase();
-                Some(format!(
-                    "Press Enter to send, Ctrl-{} for new line",
-                    newline_key
-                ))
+                Some(format!("Enter to send · Ctrl+{} newline", newline_key))
             }
         }
     }

crates/goose-cli/src/session/input.rs

@@ -398,18 +398,12 @@ fn parse_plan_command(input: String) -> Option<InputResult> {
     Some(InputResult::Plan(options))
 }
 
-/// Generates the input prompt string for the CLI interface.
-/// Returns a styled prompt with the goose face "( O)>" followed by a space.
-/// On Windows, returns plain text without ANSI styling for better compatibility.
-/// On other platforms, applies styling using ANSI escape codes.
 fn get_input_prompt_string() -> String {
-    let goose = "( O)>";
+    let goose = "🪿";
     if cfg!(target_os = "windows") {
-        // Use plain text on Windows to avoid ANSI compatibility issues
         format!("{goose} ")
     } else {
-        // On other platforms, use styled prompt with ANSI colors
-        format!("{} ", console::style(goose).cyan().bold())
+        format!("{} ", console::style(goose))
     }
 }
 
@@ -702,38 +696,14 @@ mod tests {
         let prompt = get_input_prompt_string();
 
         // Prompt should always end with a space
-        assert!(prompt.ends_with(" "));
+        assert!(prompt.ends_with(' '));
 
-        // Prompt should contain the goose face
-        assert!(prompt.contains("( O)>"));
+        // Prompt should contain the goose emoji
+        assert!(prompt.contains("🪿"));
 
-        // On Windows, prompt should be plain text without ANSI codes
         #[cfg(target_os = "windows")]
         {
-            assert_eq!(prompt, "( O)> ");
-            // Ensure no ANSI escape sequences
-            assert!(!prompt.contains("\x1b["));
-        }
-
-        // On non-Windows, prompt behavior depends on terminal capabilities
-        #[cfg(not(target_os = "windows"))]
-        {
-            // In CI environments, console crate may strip ANSI codes
-            let is_ci = std::env::var("CI").is_ok();
-
-            if is_ci {
-                // In CI, just verify basic structure - console crate handles ANSI detection
-                assert!(prompt.len() >= "( O)> ".len());
-            } else {
-                // In interactive terminals, expect styling to be applied
-                // Note: This may still vary based on terminal capabilities
-                assert!(prompt.len() >= "( O)> ".len());
-
-                // If ANSI codes are present, they should be valid
-                if prompt.contains("\x1b[") {
-                    assert!(prompt.contains("36") || prompt.contains("1"));
-                }
-            }
+            assert_eq!(prompt, "🪿 ");
         }
     }
 }

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

@@ -159,7 +159,7 @@ pub struct CliSession {
     completion_cache: Arc<std::sync::RwLock<CompletionCache>>,
     debug: bool,
     run_mode: RunMode,
-    scheduled_job_id: Option<String>, // ID of the scheduled job that triggered this session
+    scheduled_job_id: Option<String>,
     max_turns: Option<u32>,
     edit_mode: Option<EditMode>,
     retry_config: Option<RetryConfig>,
@@ -479,7 +479,6 @@ impl CliSession {
         loop {
             self.display_context_usage().await?;
 
-            // Convert conversation messages to strings for editor mode
             let conversation_strings: Vec<String> = self
                 .messages
                 .iter()
@@ -502,8 +501,9 @@ impl CliSession {
         }
 
         println!(
-            "Closing session. Session ID: {}",
-            console::style(&self.session_id).cyan()
+            "\n  {} {}",
+            console::style("●").red(),
+            console::style(format!("session closed · {}", &self.session_id)).dim()
         );
 
         Ok(())
@@ -636,10 +636,7 @@ impl CliSession {
 
                 let elapsed = start_time.elapsed();
                 let elapsed_str = format_elapsed_time(elapsed);
-                println!(
-                    "\n{}",
-                    console::style(format!("⏱️  Elapsed time: {}", elapsed_str)).dim()
-                );
+                println!("{}", console::style(format!("  ⏱ {}", elapsed_str)).dim());
             }
             RunMode::Plan => {
                 let mut plan_messages = self.messages.clone();
@@ -1152,20 +1149,10 @@ impl CliSession {
                     .collect()
             });
 
+        let interrupt_prompt = "Yes — what would you like me to do?";
+
         if !tool_requests.is_empty() {
-            // Interrupted during a tool request
-            // Create tool responses for all interrupted tool requests
-            // TODO(Douwe): if we need this, it should happen in agent reply
             let mut response_message = Message::user();
-            let last_tool_name = tool_requests
-                .last()
-                .and_then(|(_, tool_call)| {
-                    tool_call
-                        .as_ref()
-                        .ok()
-                        .map(|tool| tool.name.to_string().clone())
-                })
-                .unwrap_or_else(|| "tool".to_string());
 
             let notification = if interrupt {
                 "Interrupted by the user to make a correction".to_string()
@@ -1183,36 +1170,29 @@ impl CliSession {
                 ));
             }
             self.push_message(response_message);
-            let prompt = format!(
-                "The existing call to {} was interrupted. How would you like to proceed?",
-                last_tool_name
+            self.push_message(Message::assistant().with_text(interrupt_prompt));
+            output::render_message(
+                &Message::assistant().with_text(interrupt_prompt),
+                self.debug,
             );
-            self.push_message(Message::assistant().with_text(&prompt));
-            output::render_message(&Message::assistant().with_text(&prompt), self.debug);
-        } else {
-            // An interruption occurred outside of a tool request-response.
-            if let Some(last_msg) = self.messages.last() {
-                if last_msg.role == rmcp::model::Role::User {
-                    match last_msg.content.first() {
-                        Some(MessageContent::ToolResponse(_)) => {
-                            // Interruption occurred after a tool had completed but not assistant reply
-                            let prompt = "The tool calling loop was interrupted. How would you like to proceed?";
-                            self.push_message(Message::assistant().with_text(prompt));
-                            output::render_message(
-                                &Message::assistant().with_text(prompt),
-                                self.debug,
-                            );
-                        }
-                        Some(_) => {
-                            // A real users message
-                            self.messages.pop();
-                            let prompt = "Interrupted before the model replied and removed the last message.";
-                            output::render_message(
-                                &Message::assistant().with_text(prompt),
-                                self.debug,
-                            );
-                        }
-                        None => panic!("No content in last message"),
+        } else if let Some(last_msg) = self.messages.last() {
+            if last_msg.role == rmcp::model::Role::User {
+                match last_msg.content.first() {
+                    Some(MessageContent::ToolResponse(_)) => {
+                        self.push_message(Message::assistant().with_text(interrupt_prompt));
+                        output::render_message(
+                            &Message::assistant().with_text(interrupt_prompt),
+                            self.debug,
+                        );
+                    }
+                    Some(_) => {
+                        self.messages.pop();
+                        let assistant_msg = Message::assistant().with_text(interrupt_prompt);
+                        self.push_message(assistant_msg.clone());
+                        output::render_message(&assistant_msg, self.debug);
+                    }
+                    None => {
+                        // Empty message content — nothing to do, just continue gracefully
                     }
                 }
             }
@@ -1271,11 +1251,10 @@ impl CliSession {
             return;
         }
 
-        // Print session restored message
         println!(
-            "\n{} {} messages loaded into context.",
-            console::style("Session restored:").green().bold(),
-            console::style(self.messages.len()).green()
+            "\n  {} {}",
+            console::style("↻").cyan(),
+            console::style(format!("{} messages restored", self.messages.len())).dim()
         );
 
         // Render each message
@@ -1283,11 +1262,7 @@ impl CliSession {
             output::render_message(message, self.debug);
         }
 
-        // Add a visual separator after restored messages
-        println!(
-            "\n{}\n",
-            console::style("──────── New Messages ────────").dim()
-        );
+        println!();
     }
 
     pub async fn get_session(&self) -> Result<goose::session::Session> {

crates/goose-cli/src/session/output.rs

@@ -120,16 +120,18 @@ pub struct ThinkingIndicator {
 impl ThinkingIndicator {
     pub fn show(&mut self) {
         let spinner = cliclack::spinner();
+        let hint = style("(Ctrl+C to interrupt)").dim();
         if Config::global()
             .get_param("RANDOM_THINKING_MESSAGES")
             .unwrap_or(true)
         {
             spinner.start(format!(
-                "{}...",
-                super::thinking::get_random_thinking_message()
+                "{}...  {}",
+                super::thinking::get_random_thinking_message(),
+                hint,
             ));
         } else {
-            spinner.start("Thinking...");
+            spinner.start(format!("Thinking...  {}", hint));
         }
         self.spinner = Some(spinner);
     }
@@ -467,17 +469,15 @@ pub fn render_builtin_error(names: &str, error: &str) {
 fn render_text_editor_request(call: &CallToolRequestParams, debug: bool) {
     print_tool_header(call);
 
-    // Print path first with special formatting
     if let Some(args) = &call.arguments {
         if let Some(Value::String(path)) = args.get("path") {
             println!(
-                "{}: {}",
+                "    {} {}",
                 style("path").dim(),
-                style(shorten_path(path, debug)).green()
+                style(shorten_path(path, debug)).dim()
             );
         }
 
-        // Print other arguments normally, excluding path
         if let Some(args) = &call.arguments {
             let mut other_args = serde_json::Map::new();
             for (k, v) in args {
@@ -486,7 +486,7 @@ fn render_text_editor_request(call: &CallToolRequestParams, debug: bool) {
                 }
             }
             if !other_args.is_empty() {
-                print_params(&Some(other_args), 0, debug);
+                print_params(&Some(other_args), 1, debug);
             }
         }
     }
@@ -495,7 +495,7 @@ fn render_text_editor_request(call: &CallToolRequestParams, debug: bool) {
 
 fn render_shell_request(call: &CallToolRequestParams, debug: bool) {
     print_tool_header(call);
-    print_params(&call.arguments, 0, debug);
+    print_params(&call.arguments, 1, debug);
     println!();
 }
 
@@ -515,10 +515,11 @@ fn render_execute_code_request(call: &CallToolRequestParams, debug: bool) {
     let plural = if count == 1 { "" } else { "s" };
     println!();
     println!(
-        "─── {} tool call{} | {} ──────────────────────────",
-        style(count).cyan(),
+        "  {} {} {} tool call{}",
+        style("▸").dim(),
+        style("execute").dim(),
+        style(count).dim(),
         plural,
-        style("execute").magenta().dim()
     );
 
     for (i, node) in tool_graph.iter().filter_map(Value::as_object).enumerate() {
@@ -544,10 +545,10 @@ fn render_execute_code_request(call: &CallToolRequestParams, debug: bool) {
             format!(" (uses {})", deps.join(", "))
         };
         println!(
-            "  {}. {}: {}{}",
+            "    {}. {} {}{}",
             style(i + 1).dim(),
-            style(tool).cyan(),
-            style(desc).green(),
+            style(tool).dim(),
+            style(desc).dim(),
             style(deps_str).dim()
         );
     }
@@ -570,7 +571,7 @@ fn render_delegate_request(call: &CallToolRequestParams, debug: bool) {
 
     if let Some(args) = &call.arguments {
         if let Some(Value::String(source)) = args.get("source") {
-            println!("{}: {}", style("source").dim(), style(source).cyan());
+            println!("    {} {}", style("source").dim(), style(source).dim());
         }
 
         if let Some(Value::String(instructions)) = args.get("instructions") {
@@ -580,15 +581,15 @@ fn render_delegate_request(call: &CallToolRequestParams, debug: bool) {
                 instructions.clone()
             };
             println!(
-                "{}: {}",
+                "    {} {}",
                 style("instructions").dim(),
-                style(display).green()
+                style(display).dim()
             );
         }
 
         if let Some(Value::Object(params)) = args.get("parameters") {
-            println!("{}:", style("parameters").dim());
-            print_params(&Some(params.clone()), 1, debug);
+            println!("    {}:", style("parameters").dim());
+            print_params(&Some(params.clone()), 2, debug);
         }
 
         let skip_keys = ["source", "instructions", "parameters"];
@@ -599,7 +600,7 @@ fn render_delegate_request(call: &CallToolRequestParams, debug: bool) {
             }
         }
         if !other_args.is_empty() {
-            print_params(&Some(other_args), 0, debug);
+            print_params(&Some(other_args), 1, debug);
         }
     }
 
@@ -611,7 +612,7 @@ fn render_todo_request(call: &CallToolRequestParams, _debug: bool) {
 
     if let Some(args) = &call.arguments {
         if let Some(Value::String(content)) = args.get("content") {
-            println!("{}: {}", style("content").dim(), style(content).green());
+            println!("    {} {}", style("content").dim(), style(content).dim());
         }
     }
     println!();
@@ -619,7 +620,7 @@ fn render_todo_request(call: &CallToolRequestParams, _debug: bool) {
 
 fn render_default_request(call: &CallToolRequestParams, debug: bool) {
     print_tool_header(call);
-    print_params(&call.arguments, 0, debug);
+    print_params(&call.arguments, 1, debug);
     println!();
 }
 
@@ -660,14 +661,13 @@ pub fn render_subagent_tool_call(
         }
     }
     let tool_header = format!(
-        "─── {} ──────────────────────────",
-        style(format_subagent_tool_call_message(subagent_id, tool_name))
-            .magenta()
-            .dim()
+        "  {} {}",
+        style("▸").dim(),
+        style(format_subagent_tool_call_message(subagent_id, tool_name)).dim(),
     );
     println!();
     println!("{}", tool_header);
-    print_params(&arguments.cloned(), 0, debug);
+    print_params(&arguments.cloned(), 1, debug);
     println!();
 }
 
@@ -677,11 +677,12 @@ fn render_subagent_tool_graph(subagent_id: &str, tool_graph: &[Value]) {
     let plural = if count == 1 { "" } else { "s" };
     println!();
     println!(
-        "─── {} {} tool call{} | {} ──────────────────────────",
-        style(format!("[subagent:{}]", short_id)).cyan(),
-        style(count).cyan(),
+        "  {} {} {} {} tool call{}",
+        style("▸").dim(),
+        style(format!("[subagent:{}]", short_id)).dim(),
+        style("execute_code").dim(),
+        style(count).dim(),
         plural,
-        style("execute_code").magenta().dim()
     );
 
     for (i, node) in tool_graph.iter().filter_map(Value::as_object).enumerate() {
@@ -707,10 +708,10 @@ fn render_subagent_tool_graph(subagent_id: &str, tool_graph: &[Value]) {
             format!(" (uses {})", deps.join(", "))
         };
         println!(
-            "  {}. {}: {}{}",
+            "    {}. {} {}{}",
             style(i + 1).dim(),
-            style(tool).cyan(),
-            style(desc).green(),
+            style(tool).dim(),
+            style(desc).dim(),
             style(deps_str).dim()
         );
     }
@@ -721,11 +722,16 @@ fn render_subagent_tool_graph(subagent_id: &str, tool_graph: &[Value]) {
 
 fn print_tool_header(call: &CallToolRequestParams) {
     let (tool, extension) = split_tool_name(&call.name);
-    let tool_header = format!(
-        "─── {} | {} ──────────────────────────",
-        style(tool),
-        style(extension).magenta().dim(),
-    );
+    let tool_header = if extension.is_empty() {
+        format!("  {} {}", style("▸").dim(), style(&tool).dim())
+    } else {
+        format!(
+            "  {} {} {}",
+            style("▸").dim(),
+            style(&tool).dim(),
+            style(extension).magenta().dim(),
+        )
+    };
     println!();
     println!("{}", tool_header);
 }
@@ -889,7 +895,6 @@ fn shorten_path(path: &str, debug: bool) -> String {
     shortened.join("/")
 }
 
-// Session display functions
 pub fn display_session_info(
     resume: bool,
     provider: &str,
@@ -897,107 +902,126 @@ pub fn display_session_info(
     session_id: &Option<String>,
     provider_instance: Option<&Arc<dyn goose::providers::base::Provider>>,
 ) {
-    let start_session_msg = if resume {
-        "resuming session |"
+    let status = if resume {
+        "resuming"
     } else if session_id.is_none() {
-        "running without session |"
+        "ephemeral"
     } else {
-        "starting session |"
+        "new session"
     };
 
-    // Check if we have lead/worker mode
-    if let Some(provider_inst) = provider_instance {
+    let model_display = if let Some(provider_inst) = provider_instance {
         if let Some(lead_worker) = provider_inst.as_lead_worker() {
             let (lead_model, worker_model) = lead_worker.get_model_info();
-            println!(
-                "{} {} {} {} {} {} {}",
-                style(start_session_msg).dim(),
-                style("provider:").dim(),
-                style(provider).cyan().dim(),
-                style("lead model:").dim(),
-                style(&lead_model).cyan().dim(),
-                style("worker model:").dim(),
-                style(&worker_model).cyan().dim(),
-            );
+            format!("{} → {}", lead_model, worker_model)
         } else {
-            println!(
-                "{} {} {} {} {}",
-                style(start_session_msg).dim(),
-                style("provider:").dim(),
-                style(provider).cyan().dim(),
-                style("model:").dim(),
-                style(model).cyan().dim(),
-            );
+            model.to_string()
         }
     } else {
-        // Fallback to original behavior if no provider instance
-        println!(
-            "{} {} {} {} {}",
-            style(start_session_msg).dim(),
-            style("provider:").dim(),
-            style(provider).cyan().dim(),
-            style("model:").dim(),
-            style(model).cyan().dim(),
-        );
-    }
+        model.to_string()
+    };
+
+    println!(
+        "\n  {} {} {} {} {}",
+        style("●").green(),
+        style(status).dim(),
+        style("·").dim(),
+        style(provider).dim(),
+        style(&model_display).cyan(),
+    );
+
+    let cwd_display = std::env::current_dir()
+        .ok()
+        .map(|p| p.display().to_string())
+        .unwrap_or_else(|| "unknown".to_string());
 
     if let Some(id) = session_id {
         println!(
-            "    {} {}",
-            style("session id:").dim(),
-            style(id).cyan().dim()
+            "  {} {} {}",
+            style(" ").dim(),
+            style(id).dim(),
+            style(format!("· {}", cwd_display)).dim(),
+        );
+    } else {
+        println!(
+            "  {} {}",
+            style(" ").dim(),
+            style(format!("  {}", cwd_display)).dim(),
         );
     }
+}
 
-    println!(
-        "    {} {}",
-        style("working directory:").dim(),
-        style(std::env::current_dir().unwrap().display())
-            .cyan()
-            .dim()
-    );
+pub fn set_terminal_title() {
+    if !std::io::stdout().is_terminal() {
+        return;
+    }
+    let dir_name = std::env::current_dir()
+        .ok()
+        .and_then(|p| p.file_name().map(|n| n.to_string_lossy().into_owned()))
+        .unwrap_or_default();
+    // Sanitize: strip control characters (ESC, BEL, etc.) to prevent terminal escape injection
+    let sanitized: String = dir_name.chars().filter(|c| !c.is_control()).collect();
+    // OSC 0 sets the terminal window/tab title
+    print!("\x1b]0;🪿 {}\x07", sanitized);
+    let _ = std::io::stdout().flush();
 }
 
 pub fn display_greeting() {
-    println!("\ngoose is running! Enter your instructions, or try asking what goose can do.\n");
+    set_terminal_title();
+    println!(
+        "\n{} {}\n",
+        style("🪿 goose").bold(),
+        style("ready — type a message to get started").dim()
+    );
 }
 
-/// Display context window usage with both current and session totals
 pub fn display_context_usage(total_tokens: usize, context_limit: usize) {
     use console::style;
 
     if context_limit == 0 {
-        println!("Context: Error - context limit is zero");
+        println!(
+            "  {}",
+            style("context usage unavailable (context limit is 0)").dim()
+        );
         return;
     }
 
-    // Calculate percentage used with bounds checking
     let percentage =
         (((total_tokens as f64 / context_limit as f64) * 100.0).round() as usize).min(100);
 
-    // Create dot visualization with safety bounds
-    let dot_count = 10;
-    let filled_dots =
-        (((percentage as f64 / 100.0) * dot_count as f64).round() as usize).min(dot_count);
-    let empty_dots = dot_count - filled_dots;
+    let bar_width = 20;
+    let filled = ((percentage as f64 / 100.0) * bar_width as f64).round() as usize;
+    let empty = bar_width - filled.min(bar_width);
 
-    let filled = "●".repeat(filled_dots);
-    let empty = "○".repeat(empty_dots);
-
-    // Combine dots and apply color
-    let dots = format!("{}{}", filled, empty);
-    let colored_dots = if percentage < 50 {
-        style(dots).green()
+    let bar = format!("{}{}", "━".repeat(filled), "╌".repeat(empty));
+    let colored_bar = if percentage < 50 {
+        style(bar).green().dim()
     } else if percentage < 85 {
-        style(dots).yellow()
+        style(bar).yellow()
     } else {
-        style(dots).red()
+        style(bar).red()
     };
 
-    // Print the status line
+    fn format_tokens(n: usize) -> String {
+        if n >= 1_000_000 {
+            format!("{:.1}M", n as f64 / 1_000_000.0)
+        } else if n >= 1_000 {
+            format!("{:.0}k", n as f64 / 1_000.0)
+        } else {
+            n.to_string()
+        }
+    }
+
     println!(
-        "Context: {} {}% ({}/{} tokens)",
-        colored_dots, percentage, total_tokens, context_limit
+        "  {} {} {}",
+        colored_bar,
+        style(format!("{}%", percentage)).dim(),
+        style(format!(
+            "{}/{}",
+            format_tokens(total_tokens),
+            format_tokens(context_limit)
+        ))
+        .dim(),
     );
 }
 

scripts/test_mcp.sh

@@ -57,7 +57,7 @@ TMPFILE=$(mktemp)
 (cd "$TESTDIR" && GOOSE_PROVIDER="$TEST_PROVIDER" GOOSE_MODEL="$TEST_MODEL" \
     "$GOOSE_BIN" run --recipe recipe.yaml 2>&1) | tee "$TMPFILE"
 
-if grep -q "add | test_mcp" "$TMPFILE" && grep -q "100" "$TMPFILE"; then
+if grep -qE "(add \| test_mcp)|(▸.*add.*test_mcp)" "$TMPFILE" && grep -q "100" "$TMPFILE"; then
     echo "✓ FastMCP stderr test passed"
     RESULTS+=("✓ FastMCP stderr")
 else
@@ -73,20 +73,20 @@ TESTDIR=$(mktemp -d)
 TMPFILE=$(mktemp)
 
 (cd "$TESTDIR" && GOOSE_PROVIDER="$TEST_PROVIDER" GOOSE_MODEL="$TEST_MODEL" \
-    "$GOOSE_BIN" run --text "Use the sampleLLM tool to ask for a quote from The Great Gatsby" \
+    "$GOOSE_BIN" run --text "Use the sampleLLM tool to ask for an original short poem about the ocean" \
     --with-extension "npx -y @modelcontextprotocol/server-everything@2026.1.14" 2>&1) | tee "$TMPFILE"
 
-if grep -q "$MCP_SAMPLING_TOOL | " "$TMPFILE"; then
+if grep -qE "($MCP_SAMPLING_TOOL \| )|(▸.*$MCP_SAMPLING_TOOL)" "$TMPFILE"; then
     JUDGE_PROMPT=$(cat <<EOF
 You are a validator. You will be given a transcript of a CLI run that used an MCP tool to initiate MCP sampling.
-The MCP server requests a quote from The Great Gatsby from the model via sampling.
+The MCP server requests an original short poem about the ocean from the model via sampling.
 
 Task: Determine whether the transcript shows that the sampling request reached the model and that the output included either:
-  • A recognizable quote, paraphrase, or reference from The Great Gatsby, or
-  • A clear attempt or explanation from the model about why the quote could not be returned.
+  • A poem, verse, or creative text about the ocean or sea, or
+  • A clear attempt or explanation from the model about the poem request.
 
 If either of these conditions is true, respond PASS.
-If there is no evidence that the model attempted or returned a Gatsby-related response, respond FAIL.
+If there is no evidence that the model attempted or returned a poem-related response, respond FAIL.
 If uncertain, lean toward PASS.
 
 Output format: Respond with exactly one word on a single line:

scripts/test_providers.sh

@@ -40,7 +40,7 @@ run_test() {
       echo "failure|test content not found by model" > "$result_file"
     fi
   else
-    if ! grep -q "text_editor | developer" "$output_file"; then
+    if ! grep -qE "(text_editor \| developer)|(▸.*text_editor.*developer)" "$output_file"; then
       echo "failure|model did not use text_editor tool" > "$result_file"
     elif ! grep -q "TEST-CONTENT-ABC123" "$output_file"; then
       echo "failure|model did not return uppercased file content" > "$result_file"

scripts/test_providers_code_exec.sh

@@ -30,8 +30,9 @@ run_test() {
 
   # Verify: code_execution tool must be called
   # Matches: "execute | code_execution", "get_function_details | code_execution",
-  #           "tool call | execute", "tool calls | execute"
-  if grep -qE "(execute \| code_execution)|(get_function_details \| code_execution)|(tool calls? \| execute)" "$output_file"; then
+  #           "tool call | execute", "tool calls | execute" (old format)
+  #           "▸ execute N tool call" (new format with tool_graph)
+  if grep -qE "(execute \| code_execution)|(get_function_details \| code_execution)|(tool calls? \| execute)|(▸.*execute.*tool call)" "$output_file"; then
     echo "success|code_execution tool called" > "$result_file"
   else
     echo "failure|no code_execution tool calls found" > "$result_file"

scripts/test_subrecipes.sh

@@ -77,8 +77,8 @@ check_recipe_output() {
   local tmpfile=$1
   local mode=$2
 
-  # Check for delegate tool invocation (new format: "─── delegate |")
-  if grep -q "─── delegate" "$tmpfile"; then
+  # Check for delegate tool invocation (old: "─── delegate |", new: "▸ delegate")
+  if grep -qE "(─── delegate)|(▸.*delegate)" "$tmpfile"; then
     echo "✓ SUCCESS: Delegate tool invoked"
     RESULTS+=("✓ Delegate tool invocation ($mode)")
   else