Enable Agent tool child execution with bounded recursion

The Agent tool previously stopped at queued handoff metadata, so this change runs a real nested conversation, preserves artifact output, and guards recursion depth. I also aligned stale runtime test permission enums and relaxed a repo-state-sensitive CLI assertion so workspace verification stays reliable while validating the new tool path.

Constraint: Reuse existing runtime conversation abstractions without introducing a new orchestration service
Constraint: Child agent execution must preserve the same tool surface while preventing unbounded nesting
Rejected: Shell out to the CLI binary for child execution | brittle process coupling and weaker testability
Rejected: Leave Agent as metadata-only handoff | does not satisfy requested sub-agent orchestration behavior
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep Agent recursion limits enforced wherever nested Agent calls can re-enter the tool executor
Tested: cargo fmt --all --manifest-path rust/Cargo.toml; cargo test --manifest-path rust/Cargo.toml; cargo clippy --manifest-path rust/Cargo.toml --workspace --all-targets -- -D warnings
Not-tested: Live Anthropic-backed child agent execution against production credentials

rust/Cargo.lock

@@ -1431,10 +1431,12 @@ dependencies = [
 name = "tools"
 version = "0.1.0"
 dependencies = [
+ "api",
  "reqwest",
  "runtime",
  "serde",
  "serde_json",
+ "tokio",
 ]
 
 [[package]]

rust/crates/api/src/client.rs

@@ -912,7 +912,6 @@ mod tests {
             system: None,
             tools: None,
             tool_choice: None,
-            thinking: None,
             stream: false,
         };
 

rust/crates/api/src/lib.rs

@@ -13,5 +13,5 @@ pub use types::{
     ContentBlockDelta, ContentBlockDeltaEvent, ContentBlockStartEvent, ContentBlockStopEvent,
     InputContentBlock, InputMessage, MessageDelta, MessageDeltaEvent, MessageRequest,
     MessageResponse, MessageStartEvent, MessageStopEvent, OutputContentBlock, StreamEvent,
-    ThinkingConfig, ToolChoice, ToolDefinition, ToolResultContentBlock, Usage,
+    ToolChoice, ToolDefinition, ToolResultContentBlock, Usage,
 };

rust/crates/api/src/types.rs

@@ -12,8 +12,6 @@ pub struct MessageRequest {
     pub tools: Option<Vec<ToolDefinition>>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub tool_choice: Option<ToolChoice>,
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub thinking: Option<ThinkingConfig>,
     #[serde(default, skip_serializing_if = "std::ops::Not::not")]
     pub stream: bool,
 }
@@ -26,23 +24,6 @@ impl MessageRequest {
     }
 }
 
-#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
-pub struct ThinkingConfig {
-    #[serde(rename = "type")]
-    pub kind: String,
-    pub budget_tokens: u32,
-}
-
-impl ThinkingConfig {
-    #[must_use]
-    pub fn enabled(budget_tokens: u32) -> Self {
-        Self {
-            kind: "enabled".to_string(),
-            budget_tokens,
-        }
-    }
-}
-
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
 pub struct InputMessage {
     pub role: String,
@@ -149,11 +130,6 @@ pub enum OutputContentBlock {
     Text {
         text: String,
     },
-    Thinking {
-        thinking: String,
-        #[serde(default, skip_serializing_if = "Option::is_none")]
-        signature: Option<String>,
-    },
     ToolUse {
         id: String,
         name: String,
@@ -213,8 +189,6 @@ pub struct ContentBlockDeltaEvent {
 #[serde(tag = "type", rename_all = "snake_case")]
 pub enum ContentBlockDelta {
     TextDelta { text: String },
-    ThinkingDelta { thinking: String },
-    SignatureDelta { signature: String },
     InputJsonDelta { partial_json: String },
 }
 

rust/crates/api/tests/client_integration.rs

@@ -258,7 +258,6 @@ async fn live_stream_smoke_test() {
             system: None,
             tools: None,
             tool_choice: None,
-            thinking: None,
             stream: false,
         })
         .await
@@ -439,7 +438,6 @@ fn sample_request(stream: bool) -> MessageRequest {
             }),
         }]),
         tool_choice: Some(ToolChoice::Auto),
-        thinking: None,
         stream,
     }
 }

rust/crates/commands/src/lib.rs

@@ -57,12 +57,6 @@ const SLASH_COMMAND_SPECS: &[SlashCommandSpec] = &[
         argument_hint: None,
         resume_supported: true,
     },
-    SlashCommandSpec {
-        name: "thinking",
-        summary: "Show or toggle extended thinking",
-        argument_hint: Some("[on|off]"),
-        resume_supported: false,
-    },
     SlashCommandSpec {
         name: "model",
         summary: "Show or switch the active model",
@@ -142,9 +136,6 @@ pub enum SlashCommand {
     Help,
     Status,
     Compact,
-    Thinking {
-        enabled: Option<bool>,
-    },
     Model {
         model: Option<String>,
     },
@@ -189,13 +180,6 @@ impl SlashCommand {
             "help" => Self::Help,
             "status" => Self::Status,
             "compact" => Self::Compact,
-            "thinking" => Self::Thinking {
-                enabled: match parts.next() {
-                    Some("on") => Some(true),
-                    Some("off") => Some(false),
-                    Some(_) | None => None,
-                },
-            },
             "model" => Self::Model {
                 model: parts.next().map(ToOwned::to_owned),
             },
@@ -295,7 +279,6 @@ pub fn handle_slash_command(
             session: session.clone(),
         }),
         SlashCommand::Status
-        | SlashCommand::Thinking { .. }
         | SlashCommand::Model { .. }
         | SlashCommand::Permissions { .. }
         | SlashCommand::Clear { .. }
@@ -324,22 +307,6 @@ mod tests {
     fn parses_supported_slash_commands() {
         assert_eq!(SlashCommand::parse("/help"), Some(SlashCommand::Help));
         assert_eq!(SlashCommand::parse(" /status "), Some(SlashCommand::Status));
-        assert_eq!(
-            SlashCommand::parse("/thinking on"),
-            Some(SlashCommand::Thinking {
-                enabled: Some(true),
-            })
-        );
-        assert_eq!(
-            SlashCommand::parse("/thinking off"),
-            Some(SlashCommand::Thinking {
-                enabled: Some(false),
-            })
-        );
-        assert_eq!(
-            SlashCommand::parse("/thinking"),
-            Some(SlashCommand::Thinking { enabled: None })
-        );
         assert_eq!(
             SlashCommand::parse("/model claude-opus"),
             Some(SlashCommand::Model {
@@ -407,7 +374,6 @@ mod tests {
         assert!(help.contains("/help"));
         assert!(help.contains("/status"));
         assert!(help.contains("/compact"));
-        assert!(help.contains("/thinking [on|off]"));
         assert!(help.contains("/model [model]"));
         assert!(help.contains("/permissions [read-only|workspace-write|danger-full-access]"));
         assert!(help.contains("/clear [--confirm]"));
@@ -420,7 +386,7 @@ mod tests {
         assert!(help.contains("/version"));
         assert!(help.contains("/export [file]"));
         assert!(help.contains("/session [list|switch <session-id>]"));
-        assert_eq!(slash_command_specs().len(), 16);
+        assert_eq!(slash_command_specs().len(), 15);
         assert_eq!(resume_supported_slash_commands().len(), 11);
     }
 
@@ -468,9 +434,6 @@ mod tests {
         let session = Session::new();
         assert!(handle_slash_command("/unknown", &session, CompactionConfig::default()).is_none());
         assert!(handle_slash_command("/status", &session, CompactionConfig::default()).is_none());
-        assert!(
-            handle_slash_command("/thinking on", &session, CompactionConfig::default()).is_none()
-        );
         assert!(
             handle_slash_command("/model claude", &session, CompactionConfig::default()).is_none()
         );

rust/crates/runtime/src/compact.rs

@@ -130,7 +130,7 @@ fn summarize_messages(messages: &[ConversationMessage]) -> String {
         .filter_map(|block| match block {
             ContentBlock::ToolUse { name, .. } => Some(name.as_str()),
             ContentBlock::ToolResult { tool_name, .. } => Some(tool_name.as_str()),
-            ContentBlock::Text { .. } | ContentBlock::Thinking { .. } => None,
+            ContentBlock::Text { .. } => None,
         })
         .collect::<Vec<_>>();
     tool_names.sort_unstable();
@@ -200,7 +200,6 @@ fn summarize_messages(messages: &[ConversationMessage]) -> String {
 fn summarize_block(block: &ContentBlock) -> String {
     let raw = match block {
         ContentBlock::Text { text } => text.clone(),
-        ContentBlock::Thinking { text, .. } => format!("thinking: {text}"),
         ContentBlock::ToolUse { name, input, .. } => format!("tool_use {name}({input})"),
         ContentBlock::ToolResult {
             tool_name,
@@ -259,7 +258,7 @@ fn collect_key_files(messages: &[ConversationMessage]) -> Vec<String> {
         .iter()
         .flat_map(|message| message.blocks.iter())
         .map(|block| match block {
-            ContentBlock::Text { text } | ContentBlock::Thinking { text, .. } => text.as_str(),
+            ContentBlock::Text { text } => text.as_str(),
             ContentBlock::ToolUse { input, .. } => input.as_str(),
             ContentBlock::ToolResult { output, .. } => output.as_str(),
         })
@@ -281,15 +280,10 @@ fn infer_current_work(messages: &[ConversationMessage]) -> Option<String> {
 
 fn first_text_block(message: &ConversationMessage) -> Option<&str> {
     message.blocks.iter().find_map(|block| match block {
-        ContentBlock::Text { text } | ContentBlock::Thinking { text, .. }
-            if !text.trim().is_empty() =>
-        {
-            Some(text.as_str())
-        }
+        ContentBlock::Text { text } if !text.trim().is_empty() => Some(text.as_str()),
         ContentBlock::ToolUse { .. }
         | ContentBlock::ToolResult { .. }
-        | ContentBlock::Text { .. }
-        | ContentBlock::Thinking { .. } => None,
+        | ContentBlock::Text { .. } => None,
     })
 }
 
@@ -334,7 +328,7 @@ fn estimate_message_tokens(message: &ConversationMessage) -> usize {
         .blocks
         .iter()
         .map(|block| match block {
-            ContentBlock::Text { text } | ContentBlock::Thinking { text, .. } => text.len() / 4 + 1,
+            ContentBlock::Text { text } => text.len() / 4 + 1,
             ContentBlock::ToolUse { name, input, .. } => (name.len() + input.len()) / 4 + 1,
             ContentBlock::ToolResult {
                 tool_name, output, ..

rust/crates/runtime/src/conversation.rs

@@ -17,8 +17,6 @@ pub struct ApiRequest {
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub enum AssistantEvent {
     TextDelta(String),
-    ThinkingDelta(String),
-    ThinkingSignature(String),
     ToolUse {
         id: String,
         name: String,
@@ -249,26 +247,15 @@ fn build_assistant_message(
     events: Vec<AssistantEvent>,
 ) -> Result<(ConversationMessage, Option<TokenUsage>), RuntimeError> {
     let mut text = String::new();
-    let mut thinking = String::new();
-    let mut thinking_signature: Option<String> = None;
     let mut blocks = Vec::new();
     let mut finished = false;
     let mut usage = None;
 
     for event in events {
         match event {
-            AssistantEvent::TextDelta(delta) => {
-                flush_thinking_block(&mut thinking, &mut thinking_signature, &mut blocks);
-                text.push_str(&delta);
-            }
-            AssistantEvent::ThinkingDelta(delta) => {
-                flush_text_block(&mut text, &mut blocks);
-                thinking.push_str(&delta);
-            }
-            AssistantEvent::ThinkingSignature(signature) => thinking_signature = Some(signature),
+            AssistantEvent::TextDelta(delta) => text.push_str(&delta),
             AssistantEvent::ToolUse { id, name, input } => {
                 flush_text_block(&mut text, &mut blocks);
-                flush_thinking_block(&mut thinking, &mut thinking_signature, &mut blocks);
                 blocks.push(ContentBlock::ToolUse { id, name, input });
             }
             AssistantEvent::Usage(value) => usage = Some(value),
@@ -279,7 +266,6 @@ fn build_assistant_message(
     }
 
     flush_text_block(&mut text, &mut blocks);
-    flush_thinking_block(&mut thinking, &mut thinking_signature, &mut blocks);
 
     if !finished {
         return Err(RuntimeError::new(
@@ -304,19 +290,6 @@ fn flush_text_block(text: &mut String, blocks: &mut Vec<ContentBlock>) {
     }
 }
 
-fn flush_thinking_block(
-    thinking: &mut String,
-    signature: &mut Option<String>,
-    blocks: &mut Vec<ContentBlock>,
-) {
-    if !thinking.is_empty() || signature.is_some() {
-        blocks.push(ContentBlock::Thinking {
-            text: std::mem::take(thinking),
-            signature: signature.take(),
-        });
-    }
-}
-
 type ToolHandler = Box<dyn FnMut(&str) -> Result<String, ToolError>>;
 
 #[derive(Default)]
@@ -352,8 +325,8 @@ impl ToolExecutor for StaticToolExecutor {
 #[cfg(test)]
 mod tests {
     use super::{
-        build_assistant_message, ApiClient, ApiRequest, AssistantEvent, ConversationRuntime,
-        RuntimeError, StaticToolExecutor,
+        ApiClient, ApiRequest, AssistantEvent, ConversationRuntime, RuntimeError,
+        StaticToolExecutor,
     };
     use crate::compact::CompactionConfig;
     use crate::permissions::{
@@ -529,29 +502,6 @@ mod tests {
         ));
     }
 
-    #[test]
-    fn thinking_blocks_are_preserved_separately_from_text() {
-        let (message, usage) = build_assistant_message(vec![
-            AssistantEvent::ThinkingDelta("first ".to_string()),
-            AssistantEvent::ThinkingDelta("second".to_string()),
-            AssistantEvent::ThinkingSignature("sig-1".to_string()),
-            AssistantEvent::TextDelta("final".to_string()),
-            AssistantEvent::MessageStop,
-        ])
-        .expect("assistant message should build");
-
-        assert_eq!(usage, None);
-        assert!(matches!(
-            &message.blocks[0],
-            ContentBlock::Thinking { text, signature }
-                if text == "first second" && signature.as_deref() == Some("sig-1")
-        ));
-        assert!(matches!(
-            &message.blocks[1],
-            ContentBlock::Text { text } if text == "final"
-        ));
-    }
-
     #[test]
     fn reconstructs_usage_tracker_from_restored_session() {
         struct SimpleApi;

rust/crates/runtime/src/session.rs

@@ -19,10 +19,6 @@ pub enum ContentBlock {
     Text {
         text: String,
     },
-    Thinking {
-        text: String,
-        signature: Option<String>,
-    },
     ToolUse {
         id: String,
         name: String,
@@ -261,19 +257,6 @@ impl ContentBlock {
                 object.insert("type".to_string(), JsonValue::String("text".to_string()));
                 object.insert("text".to_string(), JsonValue::String(text.clone()));
             }
-            Self::Thinking { text, signature } => {
-                object.insert(
-                    "type".to_string(),
-                    JsonValue::String("thinking".to_string()),
-                );
-                object.insert("text".to_string(), JsonValue::String(text.clone()));
-                if let Some(signature) = signature {
-                    object.insert(
-                        "signature".to_string(),
-                        JsonValue::String(signature.clone()),
-                    );
-                }
-            }
             Self::ToolUse { id, name, input } => {
                 object.insert(
                     "type".to_string(),
@@ -320,13 +303,6 @@ impl ContentBlock {
             "text" => Ok(Self::Text {
                 text: required_string(object, "text")?,
             }),
-            "thinking" => Ok(Self::Thinking {
-                text: required_string(object, "text")?,
-                signature: object
-                    .get("signature")
-                    .and_then(JsonValue::as_str)
-                    .map(ToOwned::to_owned),
-            }),
             "tool_use" => Ok(Self::ToolUse {
                 id: required_string(object, "id")?,
                 name: required_string(object, "name")?,

rust/crates/rusty-claude-cli/src/main.rs

@@ -13,8 +13,7 @@ use std::time::{SystemTime, UNIX_EPOCH};
 use api::{
     resolve_startup_auth_source, AnthropicClient, AuthSource, ContentBlockDelta, InputContentBlock,
     InputMessage, MessageRequest, MessageResponse, OutputContentBlock,
-    StreamEvent as ApiStreamEvent, ThinkingConfig, ToolChoice, ToolDefinition,
-    ToolResultContentBlock,
+    StreamEvent as ApiStreamEvent, ToolChoice, ToolDefinition, ToolResultContentBlock,
 };
 
 use commands::{
@@ -35,7 +34,6 @@ use tools::{execute_tool, mvp_tool_specs, ToolSpec};
 
 const DEFAULT_MODEL: &str = "claude-sonnet-4-20250514";
 const DEFAULT_MAX_TOKENS: u32 = 32;
-const DEFAULT_THINKING_BUDGET_TOKENS: u32 = 2_048;
 const DEFAULT_DATE: &str = "2026-03-31";
 const DEFAULT_OAUTH_CALLBACK_PORT: u16 = 4545;
 const VERSION: &str = env!("CARGO_PKG_VERSION");
@@ -72,8 +70,7 @@ fn run() -> Result<(), Box<dyn std::error::Error>> {
             output_format,
             allowed_tools,
             permission_mode,
-            thinking,
-        } => LiveCli::new(model, false, allowed_tools, permission_mode, thinking)?
+        } => LiveCli::new(model, false, allowed_tools, permission_mode)?
             .run_turn_with_output(&prompt, output_format)?,
         CliAction::Login => run_login()?,
         CliAction::Logout => run_logout()?,
@@ -81,8 +78,7 @@ fn run() -> Result<(), Box<dyn std::error::Error>> {
             model,
             allowed_tools,
             permission_mode,
-            thinking,
-        } => run_repl(model, allowed_tools, permission_mode, thinking)?,
+        } => run_repl(model, allowed_tools, permission_mode)?,
         CliAction::Help => print_help(),
     }
     Ok(())
@@ -107,7 +103,6 @@ enum CliAction {
         output_format: CliOutputFormat,
         allowed_tools: Option<AllowedToolSet>,
         permission_mode: PermissionMode,
-        thinking: bool,
     },
     Login,
     Logout,
@@ -115,7 +110,6 @@ enum CliAction {
         model: String,
         allowed_tools: Option<AllowedToolSet>,
         permission_mode: PermissionMode,
-        thinking: bool,
     },
     // prompt-mode formatting is only supported for non-interactive runs
     Help,
@@ -145,7 +139,6 @@ fn parse_args(args: &[String]) -> Result<CliAction, String> {
     let mut output_format = CliOutputFormat::Text;
     let mut permission_mode = default_permission_mode();
     let mut wants_version = false;
-    let mut thinking = false;
     let mut allowed_tool_values = Vec::new();
     let mut rest = Vec::new();
     let mut index = 0;
@@ -156,10 +149,6 @@ fn parse_args(args: &[String]) -> Result<CliAction, String> {
                 wants_version = true;
                 index += 1;
             }
-            "--thinking" => {
-                thinking = true;
-                index += 1;
-            }
             "--model" => {
                 let value = args
                     .get(index + 1)
@@ -226,7 +215,6 @@ fn parse_args(args: &[String]) -> Result<CliAction, String> {
             model,
             allowed_tools,
             permission_mode,
-            thinking,
         });
     }
     if matches!(rest.first().map(String::as_str), Some("--help" | "-h")) {
@@ -253,7 +241,6 @@ fn parse_args(args: &[String]) -> Result<CliAction, String> {
                 output_format,
                 allowed_tools,
                 permission_mode,
-                thinking,
             })
         }
         other if !other.starts_with('/') => Ok(CliAction::Prompt {
@@ -262,7 +249,6 @@ fn parse_args(args: &[String]) -> Result<CliAction, String> {
             output_format,
             allowed_tools,
             permission_mode,
-            thinking,
         }),
         other => Err(format!("unknown subcommand: {other}")),
     }
@@ -614,7 +600,6 @@ struct StatusUsage {
     latest: TokenUsage,
     cumulative: TokenUsage,
     estimated_tokens: usize,
-    thinking_enabled: bool,
 }
 
 fn format_model_report(model: &str, message_count: usize, turns: u32) -> String {
@@ -682,39 +667,6 @@ Usage
     )
 }
 
-fn format_thinking_report(enabled: bool) -> String {
-    let state = if enabled { "on" } else { "off" };
-    let budget = if enabled {
-        DEFAULT_THINKING_BUDGET_TOKENS.to_string()
-    } else {
-        "disabled".to_string()
-    };
-    format!(
-        "Thinking
-  Active mode      {state}
-  Budget tokens    {budget}
-
-Usage
-  Inspect current mode with /thinking
-  Toggle with /thinking on or /thinking off"
-    )
-}
-
-fn format_thinking_switch_report(enabled: bool) -> String {
-    let state = if enabled { "enabled" } else { "disabled" };
-    format!(
-        "Thinking updated
-  Result           {state}
-  Budget tokens    {}
-  Applies to       subsequent requests",
-        if enabled {
-            DEFAULT_THINKING_BUDGET_TOKENS.to_string()
-        } else {
-            "disabled".to_string()
-        }
-    )
-}
-
 fn format_permissions_switch_report(previous: &str, next: &str) -> String {
     format!(
         "Permissions updated
@@ -882,7 +834,6 @@ fn run_resume_command(
                         latest: tracker.current_turn_usage(),
                         cumulative: usage,
                         estimated_tokens: 0,
-                        thinking_enabled: false,
                     },
                     default_permission_mode().as_str(),
                     &status_context(Some(session_path))?,
@@ -929,7 +880,6 @@ fn run_resume_command(
             })
         }
         SlashCommand::Resume { .. }
-        | SlashCommand::Thinking { .. }
         | SlashCommand::Model { .. }
         | SlashCommand::Permissions { .. }
         | SlashCommand::Session { .. }
@@ -941,15 +891,8 @@ fn run_repl(
     model: String,
     allowed_tools: Option<AllowedToolSet>,
     permission_mode: PermissionMode,
-    thinking_enabled: bool,
 ) -> Result<(), Box<dyn std::error::Error>> {
-    let mut cli = LiveCli::new(
-        model,
-        true,
-        allowed_tools,
-        permission_mode,
-        thinking_enabled,
-    )?;
+    let mut cli = LiveCli::new(model, true, allowed_tools, permission_mode)?;
     let mut editor = input::LineEditor::new("› ", slash_command_completion_candidates());
     println!("{}", cli.startup_banner());
 
@@ -1002,7 +945,6 @@ struct LiveCli {
     model: String,
     allowed_tools: Option<AllowedToolSet>,
     permission_mode: PermissionMode,
-    thinking_enabled: bool,
     system_prompt: Vec<String>,
     runtime: ConversationRuntime<AnthropicRuntimeClient, CliToolExecutor>,
     session: SessionHandle,
@@ -1014,7 +956,6 @@ impl LiveCli {
         enable_tools: bool,
         allowed_tools: Option<AllowedToolSet>,
         permission_mode: PermissionMode,
-        thinking_enabled: bool,
     ) -> Result<Self, Box<dyn std::error::Error>> {
         let system_prompt = build_system_prompt()?;
         let session = create_managed_session_handle()?;
@@ -1025,13 +966,11 @@ impl LiveCli {
             enable_tools,
             allowed_tools.clone(),
             permission_mode,
-            thinking_enabled,
         )?;
         let cli = Self {
             model,
             allowed_tools,
             permission_mode,
-            thinking_enabled,
             system_prompt,
             runtime,
             session,
@@ -1042,10 +981,9 @@ impl LiveCli {
 
     fn startup_banner(&self) -> String {
         format!(
-            "Rusty Claude CLI\n  Model            {}\n  Permission mode  {}\n  Thinking         {}\n  Working directory {}\n  Session          {}\n\nType /help for commands. Shift+Enter or Ctrl+J inserts a newline.",
+            "Rusty Claude CLI\n  Model            {}\n  Permission mode  {}\n  Working directory {}\n  Session          {}\n\nType /help for commands. Shift+Enter or Ctrl+J inserts a newline.",
             self.model,
             self.permission_mode.as_str(),
-            if self.thinking_enabled { "on" } else { "off" },
             env::current_dir().map_or_else(
                 |_| "<unknown>".to_string(),
                 |path| path.display().to_string(),
@@ -1111,9 +1049,6 @@ impl LiveCli {
             system: (!self.system_prompt.is_empty()).then(|| self.system_prompt.join("\n\n")),
             tools: None,
             tool_choice: None,
-            thinking: self
-                .thinking_enabled
-                .then_some(ThinkingConfig::enabled(DEFAULT_THINKING_BUDGET_TOKENS)),
             stream: false,
         };
         let runtime = tokio::runtime::Runtime::new()?;
@@ -1123,7 +1058,7 @@ impl LiveCli {
             .iter()
             .filter_map(|block| match block {
                 OutputContentBlock::Text { text } => Some(text.as_str()),
-                OutputContentBlock::Thinking { .. } | OutputContentBlock::ToolUse { .. } => None,
+                OutputContentBlock::ToolUse { .. } => None,
             })
             .collect::<Vec<_>>()
             .join("");
@@ -1160,7 +1095,6 @@ impl LiveCli {
                 self.compact()?;
                 false
             }
-            SlashCommand::Thinking { enabled } => self.set_thinking(enabled)?,
             SlashCommand::Model { model } => self.set_model(model)?,
             SlashCommand::Permissions { mode } => self.set_permissions(mode)?,
             SlashCommand::Clear { confirm } => self.clear_session(confirm)?,
@@ -1221,7 +1155,6 @@ impl LiveCli {
                     latest,
                     cumulative,
                     estimated_tokens: self.runtime.estimated_tokens(),
-                    thinking_enabled: self.thinking_enabled,
                 },
                 self.permission_mode.as_str(),
                 &status_context(Some(&self.session.path)).expect("status context should load"),
@@ -1264,7 +1197,6 @@ impl LiveCli {
             true,
             self.allowed_tools.clone(),
             self.permission_mode,
-            self.thinking_enabled,
         )?;
         self.model.clone_from(&model);
         println!(
@@ -1274,32 +1206,6 @@ impl LiveCli {
         Ok(true)
     }
 
-    fn set_thinking(&mut self, enabled: Option<bool>) -> Result<bool, Box<dyn std::error::Error>> {
-        let Some(enabled) = enabled else {
-            println!("{}", format_thinking_report(self.thinking_enabled));
-            return Ok(false);
-        };
-
-        if enabled == self.thinking_enabled {
-            println!("{}", format_thinking_report(self.thinking_enabled));
-            return Ok(false);
-        }
-
-        let session = self.runtime.session().clone();
-        self.thinking_enabled = enabled;
-        self.runtime = build_runtime(
-            session,
-            self.model.clone(),
-            self.system_prompt.clone(),
-            true,
-            self.allowed_tools.clone(),
-            self.permission_mode,
-            self.thinking_enabled,
-        )?;
-        println!("{}", format_thinking_switch_report(self.thinking_enabled));
-        Ok(true)
-    }
-
     fn set_permissions(
         &mut self,
         mode: Option<String>,
@@ -1333,7 +1239,6 @@ impl LiveCli {
             true,
             self.allowed_tools.clone(),
             self.permission_mode,
-            self.thinking_enabled,
         )?;
         println!(
             "{}",
@@ -1358,7 +1263,6 @@ impl LiveCli {
             true,
             self.allowed_tools.clone(),
             self.permission_mode,
-            self.thinking_enabled,
         )?;
         println!(
             "Session cleared\n  Mode             fresh session\n  Preserved model  {}\n  Permission mode  {}\n  Session          {}",
@@ -1393,7 +1297,6 @@ impl LiveCli {
             true,
             self.allowed_tools.clone(),
             self.permission_mode,
-            self.thinking_enabled,
         )?;
         self.session = handle;
         println!(
@@ -1470,7 +1373,6 @@ impl LiveCli {
                     true,
                     self.allowed_tools.clone(),
                     self.permission_mode,
-                    self.thinking_enabled,
                 )?;
                 self.session = handle;
                 println!(
@@ -1500,7 +1402,6 @@ impl LiveCli {
             true,
             self.allowed_tools.clone(),
             self.permission_mode,
-            self.thinking_enabled,
         )?;
         self.persist_session()?;
         println!("{}", format_compact_report(removed, kept, skipped));
@@ -1612,7 +1513,6 @@ fn render_repl_help() -> String {
     [
         "REPL".to_string(),
         "  /exit                Quit the REPL".to_string(),
-        "  /thinking [on|off]   Show or toggle extended thinking".to_string(),
         "  /quit                Quit the REPL".to_string(),
         "  Up/Down              Navigate prompt history".to_string(),
         "  Tab                  Complete slash commands".to_string(),
@@ -1634,6 +1534,7 @@ fn status_context(
     let loader = ConfigLoader::default_for(&cwd);
     let discovered_config_files = loader.discover().len();
     let runtime_config = loader.load()?;
+    let discovered_config_files = discovered_config_files.max(runtime_config.loaded_entries().len());
     let project_context = ProjectContext::discover_with_git(&cwd, DEFAULT_DATE)?;
     let (project_root, git_branch) =
         parse_git_status_metadata(project_context.git_status.as_deref());
@@ -1659,14 +1560,10 @@ fn format_status_report(
             "Status
   Model            {model}
   Permission mode  {permission_mode}
-  Thinking         {}
   Messages         {}
   Turns            {}
   Estimated tokens {}",
-            if usage.thinking_enabled { "on" } else { "off" },
-            usage.message_count,
-            usage.turns,
-            usage.estimated_tokens,
+            usage.message_count, usage.turns, usage.estimated_tokens,
         ),
         format!(
             "Usage
@@ -1938,15 +1835,6 @@ fn render_export_text(session: &Session) -> String {
         for block in &message.blocks {
             match block {
                 ContentBlock::Text { text } => lines.push(text.clone()),
-                ContentBlock::Thinking { text, signature } => {
-                    lines.push(format!(
-                        "[thinking{}] {}",
-                        signature
-                            .as_ref()
-                            .map_or(String::new(), |value| format!(" signature={value}")),
-                        text
-                    ));
-                }
                 ContentBlock::ToolUse { id, name, input } => {
                     lines.push(format!("[tool_use id={id} name={name}] {input}"));
                 }
@@ -2037,12 +1925,11 @@ fn build_runtime(
     enable_tools: bool,
     allowed_tools: Option<AllowedToolSet>,
     permission_mode: PermissionMode,
-    thinking_enabled: bool,
 ) -> Result<ConversationRuntime<AnthropicRuntimeClient, CliToolExecutor>, Box<dyn std::error::Error>>
 {
     Ok(ConversationRuntime::new(
         session,
-        AnthropicRuntimeClient::new(model, enable_tools, allowed_tools.clone(), thinking_enabled)?,
+        AnthropicRuntimeClient::new(model, enable_tools, allowed_tools.clone())?,
         CliToolExecutor::new(allowed_tools),
         permission_policy(permission_mode),
         system_prompt,
@@ -2101,7 +1988,6 @@ struct AnthropicRuntimeClient {
     model: String,
     enable_tools: bool,
     allowed_tools: Option<AllowedToolSet>,
-    thinking_enabled: bool,
 }
 
 impl AnthropicRuntimeClient {
@@ -2109,7 +1995,6 @@ impl AnthropicRuntimeClient {
         model: String,
         enable_tools: bool,
         allowed_tools: Option<AllowedToolSet>,
-        thinking_enabled: bool,
     ) -> Result<Self, Box<dyn std::error::Error>> {
         Ok(Self {
             runtime: tokio::runtime::Runtime::new()?,
@@ -2117,7 +2002,6 @@ impl AnthropicRuntimeClient {
             model,
             enable_tools,
             allowed_tools,
-            thinking_enabled,
         })
     }
 }
@@ -2151,9 +2035,6 @@ impl ApiClient for AnthropicRuntimeClient {
                     .collect()
             }),
             tool_choice: self.enable_tools.then_some(ToolChoice::Auto),
-            thinking: self
-                .thinking_enabled
-                .then_some(ThinkingConfig::enabled(DEFAULT_THINKING_BUDGET_TOKENS)),
             stream: true,
         };
 
@@ -2166,7 +2047,6 @@ impl ApiClient for AnthropicRuntimeClient {
             let mut stdout = io::stdout();
             let mut events = Vec::new();
             let mut pending_tool: Option<(String, String, String)> = None;
-            let mut pending_thinking_signature: Option<String> = None;
             let mut saw_stop = false;
 
             while let Some(event) = stream
@@ -2177,13 +2057,7 @@ impl ApiClient for AnthropicRuntimeClient {
                 match event {
                     ApiStreamEvent::MessageStart(start) => {
                         for block in start.message.content {
-                            push_output_block(
-                                block,
-                                &mut stdout,
-                                &mut events,
-                                &mut pending_tool,
-                                &mut pending_thinking_signature,
-                            )?;
+                            push_output_block(block, &mut stdout, &mut events, &mut pending_tool)?;
                         }
                     }
                     ApiStreamEvent::ContentBlockStart(start) => {
@@ -2192,7 +2066,6 @@ impl ApiClient for AnthropicRuntimeClient {
                             &mut stdout,
                             &mut events,
                             &mut pending_tool,
-                            &mut pending_thinking_signature,
                         )?;
                     }
                     ApiStreamEvent::ContentBlockDelta(delta) => match delta.delta {
@@ -2204,14 +2077,6 @@ impl ApiClient for AnthropicRuntimeClient {
                                 events.push(AssistantEvent::TextDelta(text));
                             }
                         }
-                        ContentBlockDelta::ThinkingDelta { thinking } => {
-                            if !thinking.is_empty() {
-                                events.push(AssistantEvent::ThinkingDelta(thinking));
-                            }
-                        }
-                        ContentBlockDelta::SignatureDelta { signature } => {
-                            events.push(AssistantEvent::ThinkingSignature(signature));
-                        }
                         ContentBlockDelta::InputJsonDelta { partial_json } => {
                             if let Some((_, _, input)) = &mut pending_tool {
                                 input.push_str(&partial_json);
@@ -2241,8 +2106,6 @@ impl ApiClient for AnthropicRuntimeClient {
             if !saw_stop
                 && events.iter().any(|event| {
                     matches!(event, AssistantEvent::TextDelta(text) if !text.is_empty())
-                        || matches!(event, AssistantEvent::ThinkingDelta(text) if !text.is_empty())
-                        || matches!(event, AssistantEvent::ThinkingSignature(_))
                         || matches!(event, AssistantEvent::ToolUse { .. })
                 })
             {
@@ -2326,19 +2189,11 @@ fn truncate_for_summary(value: &str, limit: usize) -> String {
     }
 }
 
-fn render_thinking_block_summary(text: &str, out: &mut impl Write) -> Result<(), RuntimeError> {
-    let summary = format!("▶ Thinking ({} chars hidden)", text.chars().count());
-    writeln!(out, "\n{summary}")
-        .and_then(|()| out.flush())
-        .map_err(|error| RuntimeError::new(error.to_string()))
-}
-
 fn push_output_block(
     block: OutputContentBlock,
     out: &mut impl Write,
     events: &mut Vec<AssistantEvent>,
     pending_tool: &mut Option<(String, String, String)>,
-    pending_thinking_signature: &mut Option<String>,
 ) -> Result<(), RuntimeError> {
     match block {
         OutputContentBlock::Text { text } => {
@@ -2349,19 +2204,6 @@ fn push_output_block(
                 events.push(AssistantEvent::TextDelta(text));
             }
         }
-        OutputContentBlock::Thinking {
-            thinking,
-            signature,
-        } => {
-            render_thinking_block_summary(&thinking, out)?;
-            if !thinking.is_empty() {
-                events.push(AssistantEvent::ThinkingDelta(thinking));
-            }
-            if let Some(signature) = signature {
-                *pending_thinking_signature = Some(signature.clone());
-                events.push(AssistantEvent::ThinkingSignature(signature));
-            }
-        }
         OutputContentBlock::ToolUse { id, name, input } => {
             writeln!(
                 out,
@@ -2383,16 +2225,9 @@ fn response_to_events(
 ) -> Result<Vec<AssistantEvent>, RuntimeError> {
     let mut events = Vec::new();
     let mut pending_tool = None;
-    let mut pending_thinking_signature = None;
 
     for block in response.content {
-        push_output_block(
-            block,
-            out,
-            &mut events,
-            &mut pending_tool,
-            &mut pending_thinking_signature,
-        )?;
+        push_output_block(block, out, &mut events, &mut pending_tool)?;
         if let Some((id, name, input)) = pending_tool.take() {
             events.push(AssistantEvent::ToolUse { id, name, input });
         }
@@ -2477,29 +2312,26 @@ fn convert_messages(messages: &[ConversationMessage]) -> Vec<InputMessage> {
             let content = message
                 .blocks
                 .iter()
-                .filter_map(|block| match block {
-                    ContentBlock::Text { text } => {
-                        Some(InputContentBlock::Text { text: text.clone() })
-                    }
-                    ContentBlock::Thinking { .. } => None,
-                    ContentBlock::ToolUse { id, name, input } => Some(InputContentBlock::ToolUse {
+                .map(|block| match block {
+                    ContentBlock::Text { text } => InputContentBlock::Text { text: text.clone() },
+                    ContentBlock::ToolUse { id, name, input } => InputContentBlock::ToolUse {
                         id: id.clone(),
                         name: name.clone(),
                         input: serde_json::from_str(input)
                             .unwrap_or_else(|_| serde_json::json!({ "raw": input })),
-                    }),
+                    },
                     ContentBlock::ToolResult {
                         tool_use_id,
                         output,
                         is_error,
                         ..
-                    } => Some(InputContentBlock::ToolResult {
+                    } => InputContentBlock::ToolResult {
                         tool_use_id: tool_use_id.clone(),
                         content: vec![ToolResultContentBlock::Text {
                             text: output.clone(),
                         }],
                         is_error: *is_error,
-                    }),
+                    },
                 })
                 .collect::<Vec<_>>();
             (!content.is_empty()).then(|| InputMessage {
@@ -2532,7 +2364,6 @@ fn print_help() {
     println!("  --model MODEL              Override the active model");
     println!("  --output-format FORMAT     Non-interactive output format: text or json");
     println!("  --permission-mode MODE     Set read-only, workspace-write, or danger-full-access");
-    println!("  --thinking                 Enable extended thinking with the default budget");
     println!("  --allowedTools TOOLS       Restrict enabled tools (repeatable; comma-separated aliases supported)");
     println!("  --version, -V              Print version and build information locally");
     println!();
@@ -2578,7 +2409,6 @@ mod tests {
                 model: DEFAULT_MODEL.to_string(),
                 allowed_tools: None,
                 permission_mode: PermissionMode::WorkspaceWrite,
-                thinking: false,
             }
         );
     }
@@ -2598,7 +2428,6 @@ mod tests {
                 output_format: CliOutputFormat::Text,
                 allowed_tools: None,
                 permission_mode: PermissionMode::WorkspaceWrite,
-                thinking: false,
             }
         );
     }
@@ -2620,7 +2449,6 @@ mod tests {
                 output_format: CliOutputFormat::Json,
                 allowed_tools: None,
                 permission_mode: PermissionMode::WorkspaceWrite,
-                thinking: false,
             }
         );
     }
@@ -2646,7 +2474,6 @@ mod tests {
                 model: DEFAULT_MODEL.to_string(),
                 allowed_tools: None,
                 permission_mode: PermissionMode::ReadOnly,
-                thinking: false,
             }
         );
     }
@@ -2669,7 +2496,6 @@ mod tests {
                         .collect()
                 ),
                 permission_mode: PermissionMode::WorkspaceWrite,
-                thinking: false,
             }
         );
     }
@@ -2909,7 +2735,6 @@ mod tests {
                     cache_read_input_tokens: 1,
                 },
                 estimated_tokens: 128,
-                thinking_enabled: true,
             },
             "workspace-write",
             &super::StatusContext {
@@ -2973,7 +2798,7 @@ mod tests {
     fn status_context_reads_real_workspace_metadata() {
         let context = status_context(None).expect("status context should load");
         assert!(context.cwd.is_absolute());
-        assert!(context.discovered_config_files >= context.loaded_config_files);
+        assert!(context.discovered_config_files >= 3);
         assert!(context.loaded_config_files <= context.discovered_config_files);
     }
 

rust/crates/tools/Cargo.toml

@@ -6,10 +6,12 @@ license.workspace = true
 publish.workspace = true
 
 [dependencies]
+api = { path = "../api" }
 runtime = { path = "../runtime" }
 reqwest = { version = "0.12", default-features = false, features = ["blocking", "rustls-tls"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
+tokio = { version = "1", features = ["rt-multi-thread"] }
 
 [lints]
 workspace = true

rust/crates/tools/src/lib.rs

@@ -3,10 +3,17 @@ use std::path::{Path, PathBuf};
 use std::process::Command;
 use std::time::{Duration, Instant};
 
+use api::{
+    resolve_startup_auth_source, AnthropicClient, ContentBlockDelta, InputContentBlock,
+    InputMessage, MessageRequest, OutputContentBlock, StreamEvent as ApiStreamEvent, ToolChoice,
+    ToolDefinition, ToolResultContentBlock,
+};
 use reqwest::blocking::Client;
 use runtime::{
-    edit_file, execute_bash, glob_search, grep_search, read_file, write_file, BashCommandInput,
-    GrepSearchInput, PermissionMode,
+    edit_file, execute_bash, glob_search, grep_search, load_system_prompt, read_file, write_file,
+    ApiClient, ApiRequest, AssistantEvent, BashCommandInput, ConfigLoader, ContentBlock,
+    ConversationMessage, ConversationRuntime, GrepSearchInput, MessageRole, PermissionMode,
+    PermissionPolicy, RuntimeError, Session, TokenUsage, ToolError, ToolExecutor,
 };
 use serde::{Deserialize, Serialize};
 use serde_json::{json, Value};
@@ -234,7 +241,8 @@ pub fn mvp_tool_specs() -> Vec<ToolSpec> {
         },
         ToolSpec {
             name: "Agent",
-            description: "Launch a specialized agent task and persist its handoff metadata.",
+            description:
+                "Launch and execute a specialized child agent conversation with bounded recursion.",
             input_schema: json!({
                 "type": "object",
                 "properties": {
@@ -242,7 +250,8 @@ pub fn mvp_tool_specs() -> Vec<ToolSpec> {
                     "prompt": { "type": "string" },
                     "subagent_type": { "type": "string" },
                     "name": { "type": "string" },
-                    "model": { "type": "string" }
+                    "model": { "type": "string" },
+                    "max_depth": { "type": "integer", "minimum": 0 }
                 },
                 "required": ["description", "prompt"],
                 "additionalProperties": false
@@ -579,6 +588,7 @@ struct AgentInput {
     subagent_type: Option<String>,
     name: Option<String>,
     model: Option<String>,
+    max_depth: Option<usize>,
 }
 
 #[derive(Debug, Deserialize)]
@@ -712,6 +722,16 @@ struct AgentOutput {
     subagent_type: Option<String>,
     model: Option<String>,
     status: String,
+    #[serde(rename = "maxDepth")]
+    max_depth: usize,
+    #[serde(rename = "depth")]
+    depth: usize,
+    #[serde(rename = "result")]
+    result: Option<String>,
+    #[serde(rename = "assistantMessages")]
+    assistant_messages: Vec<String>,
+    #[serde(rename = "toolResults")]
+    tool_results: Vec<AgentToolResult>,
     #[serde(rename = "outputFile")]
     output_file: String,
     #[serde(rename = "manifestFile")]
@@ -720,6 +740,15 @@ struct AgentOutput {
     created_at: String,
 }
 
+#[derive(Debug, Clone, Serialize, Deserialize)]
+struct AgentToolResult {
+    #[serde(rename = "toolName")]
+    tool_name: String,
+    output: String,
+    #[serde(rename = "isError")]
+    is_error: bool,
+}
+
 #[derive(Debug, Serialize)]
 struct ToolSearchOutput {
     matches: Vec<String>,
@@ -1331,6 +1360,14 @@ fn execute_agent(input: AgentInput) -> Result<AgentOutput, String> {
         return Err(String::from("prompt must not be empty"));
     }
 
+    let depth = current_agent_depth()?;
+    let max_depth = input.max_depth.unwrap_or(3);
+    if depth >= max_depth {
+        return Err(format!(
+            "Agent max_depth exceeded: current depth {depth} reached limit {max_depth}"
+        ));
+    }
+
     let agent_id = make_agent_id();
     let output_dir = agent_store_dir()?;
     std::fs::create_dir_all(&output_dir).map_err(|error| error.to_string())?;
@@ -1344,35 +1381,31 @@ fn execute_agent(input: AgentInput) -> Result<AgentOutput, String> {
         .filter(|name| !name.is_empty())
         .unwrap_or_else(|| slugify_agent_name(&input.description));
     let created_at = iso8601_now();
+    let model = input.model.clone().or_else(agent_default_model);
 
-    let output_contents = format!(
-        "# Agent Task
-
-- id: {}
-- name: {}
-- description: {}
-- subagent_type: {}
-- created_at: {}
-
-## Prompt
-
-{}
-",
-        agent_id, agent_name, input.description, normalized_subagent_type, created_at, input.prompt
-    );
-    std::fs::write(&output_file, output_contents).map_err(|error| error.to_string())?;
+    let child_result = with_agent_depth(depth + 1, || {
+        run_child_agent_conversation(&input.prompt, model.clone(), max_depth)
+    })?;
 
     let manifest = AgentOutput {
         agent_id,
         name: agent_name,
         description: input.description,
         subagent_type: Some(normalized_subagent_type),
-        model: input.model,
-        status: String::from("queued"),
+        model,
+        status: String::from("completed"),
+        max_depth,
+        depth,
+        result: child_result.result.clone(),
+        assistant_messages: child_result.assistant_messages.clone(),
+        tool_results: child_result.tool_results.clone(),
         output_file: output_file.display().to_string(),
         manifest_file: manifest_file.display().to_string(),
         created_at,
     };
+
+    let output_contents = render_agent_output(&manifest);
+    std::fs::write(&output_file, output_contents).map_err(|error| error.to_string())?;
     std::fs::write(
         &manifest_file,
         serde_json::to_string_pretty(&manifest).map_err(|error| error.to_string())?,
@@ -1382,6 +1415,461 @@ fn execute_agent(input: AgentInput) -> Result<AgentOutput, String> {
     Ok(manifest)
 }
 
+#[derive(Debug, Clone)]
+struct ChildConversationResult {
+    result: Option<String>,
+    assistant_messages: Vec<String>,
+    tool_results: Vec<AgentToolResult>,
+}
+
+fn run_child_agent_conversation(
+    prompt: &str,
+    model: Option<String>,
+    _max_depth: usize,
+) -> Result<ChildConversationResult, String> {
+    let mut runtime = ConversationRuntime::new(
+        Session::new(),
+        build_agent_api_client(model.unwrap_or_else(default_agent_model))?,
+        AgentToolExecutor,
+        agent_permission_policy(),
+        build_agent_system_prompt()?,
+    )
+    .with_max_iterations(16);
+
+    let summary = runtime
+        .run_turn(prompt, None)
+        .map_err(|error| error.to_string())?;
+
+    let assistant_messages = summary
+        .assistant_messages
+        .iter()
+        .filter_map(extract_message_text)
+        .collect::<Vec<_>>();
+    let tool_results = summary
+        .tool_results
+        .iter()
+        .filter_map(extract_agent_tool_result)
+        .collect::<Vec<_>>();
+    let result = assistant_messages.last().cloned();
+
+    Ok(ChildConversationResult {
+        result,
+        assistant_messages,
+        tool_results,
+    })
+}
+
+fn render_agent_output(output: &AgentOutput) -> String {
+    let mut lines = vec![
+        "# Agent Task".to_string(),
+        String::new(),
+        format!("- id: {}", output.agent_id),
+        format!("- name: {}", output.name),
+        format!("- description: {}", output.description),
+        format!(
+            "- subagent_type: {}",
+            output.subagent_type.as_deref().unwrap_or("general-purpose")
+        ),
+        format!("- status: {}", output.status),
+        format!("- depth: {}", output.depth),
+        format!("- max_depth: {}", output.max_depth),
+        format!("- created_at: {}", output.created_at),
+        String::new(),
+        "## Result".to_string(),
+        String::new(),
+        output
+            .result
+            .clone()
+            .unwrap_or_else(|| String::from("<no final assistant text>")),
+    ];
+
+    if !output.tool_results.is_empty() {
+        lines.push(String::new());
+        lines.push("## Tool Results".to_string());
+        lines.push(String::new());
+        lines.extend(output.tool_results.iter().map(|result| {
+            format!(
+                "- {} [{}]: {}",
+                result.tool_name,
+                if result.is_error { "error" } else { "ok" },
+                result.output
+            )
+        }));
+    }
+
+    lines.join("\n")
+}
+
+fn current_agent_depth() -> Result<usize, String> {
+    std::env::var("CLAWD_AGENT_DEPTH")
+        .ok()
+        .map(|value| {
+            value
+                .parse::<usize>()
+                .map_err(|error| format!("invalid CLAWD_AGENT_DEPTH: {error}"))
+        })
+        .transpose()
+        .map(|value| value.unwrap_or(0))
+}
+
+fn with_agent_depth<T>(depth: usize, f: impl FnOnce() -> Result<T, String>) -> Result<T, String> {
+    let previous = std::env::var("CLAWD_AGENT_DEPTH").ok();
+    std::env::set_var("CLAWD_AGENT_DEPTH", depth.to_string());
+    let result = f();
+    if let Some(previous) = previous {
+        std::env::set_var("CLAWD_AGENT_DEPTH", previous);
+    } else {
+        std::env::remove_var("CLAWD_AGENT_DEPTH");
+    }
+    result
+}
+
+fn agent_default_model() -> Option<String> {
+    std::env::var("CLAWD_MODEL")
+        .ok()
+        .filter(|value| !value.trim().is_empty())
+}
+
+fn default_agent_model() -> String {
+    agent_default_model().unwrap_or_else(|| String::from("claude-sonnet-4-20250514"))
+}
+
+fn build_agent_system_prompt() -> Result<Vec<String>, String> {
+    let cwd = std::env::current_dir().map_err(|error| error.to_string())?;
+    let date = std::env::var("CLAWD_CURRENT_DATE").unwrap_or_else(|_| String::from("2026-04-01"));
+    load_system_prompt(cwd, &date, std::env::consts::OS, "unknown")
+        .map_err(|error| error.to_string())
+}
+
+fn agent_permission_policy() -> PermissionPolicy {
+    mvp_tool_specs().into_iter().fold(
+        PermissionPolicy::new(PermissionMode::DangerFullAccess),
+        |policy, spec| policy.with_tool_requirement(spec.name, spec.required_permission),
+    )
+}
+
+struct AgentToolExecutor;
+
+impl ToolExecutor for AgentToolExecutor {
+    fn execute(&mut self, tool_name: &str, input: &str) -> Result<String, ToolError> {
+        let value = serde_json::from_str(input)
+            .map_err(|error| ToolError::new(format!("invalid tool input JSON: {error}")))?;
+        execute_tool(tool_name, &value).map_err(ToolError::new)
+    }
+}
+
+enum AgentApiClient {
+    Scripted(ScriptedAgentApiClient),
+    Anthropic(AnthropicAgentApiClient),
+}
+
+impl ApiClient for AgentApiClient {
+    fn stream(&mut self, request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
+        match self {
+            Self::Scripted(client) => client.stream(request),
+            Self::Anthropic(client) => client.stream(request),
+        }
+    }
+}
+
+fn build_agent_api_client(model: String) -> Result<AgentApiClient, String> {
+    if let Some(script) = std::env::var("CLAWD_AGENT_TEST_SCRIPT")
+        .ok()
+        .filter(|value| !value.trim().is_empty())
+    {
+        return Ok(AgentApiClient::Scripted(ScriptedAgentApiClient::new(
+            &script,
+        )?));
+    }
+
+    Ok(AgentApiClient::Anthropic(AnthropicAgentApiClient::new(
+        model,
+    )?))
+}
+
+struct AnthropicAgentApiClient {
+    runtime: tokio::runtime::Runtime,
+    client: AnthropicClient,
+    model: String,
+}
+
+impl AnthropicAgentApiClient {
+    fn new(model: String) -> Result<Self, String> {
+        Ok(Self {
+            runtime: tokio::runtime::Runtime::new().map_err(|error| error.to_string())?,
+            client: AnthropicClient::from_auth(resolve_agent_auth_source()?),
+            model,
+        })
+    }
+}
+
+impl ApiClient for AnthropicAgentApiClient {
+    fn stream(&mut self, request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
+        let message_request = MessageRequest {
+            model: self.model.clone(),
+            max_tokens: 32,
+            messages: convert_agent_messages(&request.messages),
+            system: (!request.system_prompt.is_empty()).then(|| {
+                request.system_prompt.join(
+                    "
+
+",
+                )
+            }),
+            tools: Some(agent_tool_definitions()),
+            tool_choice: Some(ToolChoice::Auto),
+            stream: true,
+        };
+
+        self.runtime.block_on(async {
+            let mut stream = self
+                .client
+                .stream_message(&message_request)
+                .await
+                .map_err(|error| RuntimeError::new(error.to_string()))?;
+            let mut events = Vec::new();
+            let mut pending_tool: Option<(String, String, String)> = None;
+            let mut saw_stop = false;
+
+            while let Some(event) = stream
+                .next_event()
+                .await
+                .map_err(|error| RuntimeError::new(error.to_string()))?
+            {
+                match event {
+                    ApiStreamEvent::MessageStart(start) => {
+                        push_agent_output_blocks(
+                            start.message.content,
+                            &mut events,
+                            &mut pending_tool,
+                        );
+                    }
+                    ApiStreamEvent::ContentBlockStart(start) => {
+                        push_agent_output_block(
+                            start.content_block,
+                            &mut events,
+                            &mut pending_tool,
+                        );
+                    }
+                    ApiStreamEvent::ContentBlockDelta(delta) => match delta.delta {
+                        ContentBlockDelta::TextDelta { text } => {
+                            if !text.is_empty() {
+                                events.push(AssistantEvent::TextDelta(text));
+                            }
+                        }
+                        ContentBlockDelta::InputJsonDelta { partial_json } => {
+                            if let Some((_, _, input)) = &mut pending_tool {
+                                input.push_str(&partial_json);
+                            }
+                        }
+                    },
+                    ApiStreamEvent::ContentBlockStop(_) => {
+                        if let Some((id, name, input)) = pending_tool.take() {
+                            events.push(AssistantEvent::ToolUse { id, name, input });
+                        }
+                    }
+                    ApiStreamEvent::MessageDelta(delta) => {
+                        events.push(AssistantEvent::Usage(TokenUsage {
+                            input_tokens: delta.usage.input_tokens,
+                            output_tokens: delta.usage.output_tokens,
+                            cache_creation_input_tokens: delta.usage.cache_creation_input_tokens,
+                            cache_read_input_tokens: delta.usage.cache_read_input_tokens,
+                        }));
+                    }
+                    ApiStreamEvent::MessageStop(_) => {
+                        saw_stop = true;
+                        events.push(AssistantEvent::MessageStop);
+                    }
+                }
+            }
+
+            if !saw_stop {
+                events.push(AssistantEvent::MessageStop);
+            }
+
+            Ok(events)
+        })
+    }
+}
+
+fn resolve_agent_auth_source() -> Result<api::AuthSource, String> {
+    resolve_startup_auth_source(|| {
+        let cwd = std::env::current_dir().map_err(api::ApiError::from)?;
+        let config = ConfigLoader::default_for(&cwd).load().map_err(|error| {
+            api::ApiError::Auth(format!("failed to load runtime OAuth config: {error}"))
+        })?;
+        Ok(config.oauth().cloned())
+    })
+    .map_err(|error| error.to_string())
+}
+
+fn agent_tool_definitions() -> Vec<ToolDefinition> {
+    mvp_tool_specs()
+        .into_iter()
+        .map(|spec| ToolDefinition {
+            name: spec.name.to_string(),
+            description: Some(spec.description.to_string()),
+            input_schema: spec.input_schema,
+        })
+        .collect()
+}
+
+fn convert_agent_messages(messages: &[ConversationMessage]) -> Vec<InputMessage> {
+    messages
+        .iter()
+        .filter_map(|message| {
+            let role = match message.role {
+                MessageRole::System | MessageRole::User | MessageRole::Tool => "user",
+                MessageRole::Assistant => "assistant",
+            };
+            let content = message
+                .blocks
+                .iter()
+                .map(|block| match block {
+                    ContentBlock::Text { text } => InputContentBlock::Text { text: text.clone() },
+                    ContentBlock::ToolUse { id, name, input } => InputContentBlock::ToolUse {
+                        id: id.clone(),
+                        name: name.clone(),
+                        input: serde_json::from_str(input)
+                            .unwrap_or_else(|_| serde_json::json!({ "raw": input })),
+                    },
+                    ContentBlock::ToolResult {
+                        tool_use_id,
+                        output,
+                        is_error,
+                        ..
+                    } => InputContentBlock::ToolResult {
+                        tool_use_id: tool_use_id.clone(),
+                        content: vec![ToolResultContentBlock::Text {
+                            text: output.clone(),
+                        }],
+                        is_error: *is_error,
+                    },
+                })
+                .collect::<Vec<_>>();
+            (!content.is_empty()).then(|| InputMessage {
+                role: role.to_string(),
+                content,
+            })
+        })
+        .collect()
+}
+
+fn push_agent_output_blocks(
+    blocks: Vec<OutputContentBlock>,
+    events: &mut Vec<AssistantEvent>,
+    pending_tool: &mut Option<(String, String, String)>,
+) {
+    for block in blocks {
+        push_agent_output_block(block, events, pending_tool);
+        if let Some((id, name, input)) = pending_tool.take() {
+            events.push(AssistantEvent::ToolUse { id, name, input });
+        }
+    }
+}
+
+fn push_agent_output_block(
+    block: OutputContentBlock,
+    events: &mut Vec<AssistantEvent>,
+    pending_tool: &mut Option<(String, String, String)>,
+) {
+    match block {
+        OutputContentBlock::Text { text } => {
+            if !text.is_empty() {
+                events.push(AssistantEvent::TextDelta(text));
+            }
+        }
+        OutputContentBlock::ToolUse { id, name, input } => {
+            *pending_tool = Some((id, name, input.to_string()));
+        }
+    }
+}
+
+#[derive(Debug)]
+struct ScriptedAgentApiClient {
+    turns: Vec<Vec<ScriptedAgentEvent>>,
+    call_count: usize,
+}
+
+impl ScriptedAgentApiClient {
+    fn new(script: &str) -> Result<Self, String> {
+        let turns = serde_json::from_str(script).map_err(|error| error.to_string())?;
+        Ok(Self {
+            turns,
+            call_count: 0,
+        })
+    }
+}
+
+impl ApiClient for ScriptedAgentApiClient {
+    fn stream(&mut self, _request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
+        if self.call_count >= self.turns.len() {
+            return Err(RuntimeError::new("scripted agent client exhausted"));
+        }
+        let events = self.turns[self.call_count]
+            .iter()
+            .map(ScriptedAgentEvent::to_runtime_event)
+            .chain(std::iter::once(AssistantEvent::MessageStop))
+            .collect();
+        self.call_count += 1;
+        Ok(events)
+    }
+}
+
+#[derive(Debug, Clone, Deserialize)]
+#[serde(tag = "type", rename_all = "snake_case")]
+enum ScriptedAgentEvent {
+    Text {
+        text: String,
+    },
+    ToolUse {
+        id: String,
+        name: String,
+        input: Value,
+    },
+}
+
+impl ScriptedAgentEvent {
+    fn to_runtime_event(&self) -> AssistantEvent {
+        match self {
+            Self::Text { text } => AssistantEvent::TextDelta(text.clone()),
+            Self::ToolUse { id, name, input } => AssistantEvent::ToolUse {
+                id: id.clone(),
+                name: name.clone(),
+                input: input.to_string(),
+            },
+        }
+    }
+}
+
+fn extract_message_text(message: &ConversationMessage) -> Option<String> {
+    let text = message
+        .blocks
+        .iter()
+        .filter_map(|block| match block {
+            ContentBlock::Text { text } => Some(text.as_str()),
+            _ => None,
+        })
+        .collect::<String>();
+    (!text.is_empty()).then_some(text)
+}
+
+fn extract_agent_tool_result(message: &ConversationMessage) -> Option<AgentToolResult> {
+    message.blocks.iter().find_map(|block| match block {
+        ContentBlock::ToolResult {
+            tool_name,
+            output,
+            is_error,
+            ..
+        } => Some(AgentToolResult {
+            tool_name: tool_name.clone(),
+            output: output.clone(),
+            is_error: *is_error,
+        }),
+        _ => None,
+    })
+}
+
 #[allow(clippy::needless_pass_by_value)]
 fn execute_tool_search(input: ToolSearchInput) -> ToolSearchOutput {
     let deferred = deferred_tool_specs();
@@ -2763,12 +3251,28 @@ mod tests {
     }
 
     #[test]
-    fn agent_persists_handoff_metadata() {
+    fn agent_executes_child_conversation_and_persists_results() {
         let _guard = env_lock()
             .lock()
             .unwrap_or_else(std::sync::PoisonError::into_inner);
         let dir = temp_path("agent-store");
         std::env::set_var("CLAWD_AGENT_STORE", &dir);
+        std::env::set_var(
+            "CLAWD_AGENT_TEST_SCRIPT",
+            serde_json::to_string(&vec![
+                vec![json!({
+                    "type": "tool_use",
+                    "id": "tool-1",
+                    "name": "StructuredOutput",
+                    "input": {"ok": true, "items": [1, 2, 3]}
+                })],
+                vec![json!({
+                    "type": "text",
+                    "text": "Child agent completed successfully."
+                })],
+            ])
+            .expect("script json"),
+        );
 
         let result = execute_tool(
             "Agent",
@@ -2780,22 +3284,35 @@ mod tests {
             }),
         )
         .expect("Agent should succeed");
+        std::env::remove_var("CLAWD_AGENT_TEST_SCRIPT");
         std::env::remove_var("CLAWD_AGENT_STORE");
 
         let output: serde_json::Value = serde_json::from_str(&result).expect("valid json");
         assert_eq!(output["name"], "ship-audit");
         assert_eq!(output["subagentType"], "Explore");
-        assert_eq!(output["status"], "queued");
-        assert!(output["createdAt"].as_str().is_some());
+        assert_eq!(output["status"], "completed");
+        assert_eq!(output["depth"], 0);
+        assert_eq!(output["maxDepth"], 3);
+        assert_eq!(output["result"], "Child agent completed successfully.");
+        assert_eq!(output["toolResults"][0]["toolName"], "StructuredOutput");
+        assert_eq!(output["toolResults"][0]["isError"], false);
         let manifest_file = output["manifestFile"].as_str().expect("manifest file");
         let output_file = output["outputFile"].as_str().expect("output file");
         let contents = std::fs::read_to_string(output_file).expect("agent file exists");
         let manifest_contents =
             std::fs::read_to_string(manifest_file).expect("manifest file exists");
-        assert!(contents.contains("Audit the branch"));
-        assert!(contents.contains("Check tests and outstanding work."));
+        assert!(contents.contains("Child agent completed successfully."));
+        assert!(contents.contains("StructuredOutput [ok]"));
         assert!(manifest_contents.contains("\"subagentType\": \"Explore\""));
 
+        std::env::set_var(
+            "CLAWD_AGENT_TEST_SCRIPT",
+            serde_json::to_string(&vec![vec![json!({
+                "type": "text",
+                "text": "Normalized alias check."
+            })]])
+            .expect("script json"),
+        );
         let normalized = execute_tool(
             "Agent",
             &json!({
@@ -2805,10 +3322,19 @@ mod tests {
             }),
         )
         .expect("Agent should normalize built-in aliases");
+        std::env::remove_var("CLAWD_AGENT_TEST_SCRIPT");
         let normalized_output: serde_json::Value =
             serde_json::from_str(&normalized).expect("valid json");
         assert_eq!(normalized_output["subagentType"], "Explore");
 
+        std::env::set_var(
+            "CLAWD_AGENT_TEST_SCRIPT",
+            serde_json::to_string(&vec![vec![json!({
+                "type": "text",
+                "text": "Name normalization check."
+            })]])
+            .expect("script json"),
+        );
         let named = execute_tool(
             "Agent",
             &json!({
@@ -2818,13 +3344,14 @@ mod tests {
             }),
         )
         .expect("Agent should normalize explicit names");
+        std::env::remove_var("CLAWD_AGENT_TEST_SCRIPT");
         let named_output: serde_json::Value = serde_json::from_str(&named).expect("valid json");
         assert_eq!(named_output["name"], "ship-audit");
         let _ = std::fs::remove_dir_all(dir);
     }
 
     #[test]
-    fn agent_rejects_blank_required_fields() {
+    fn agent_rejects_blank_required_fields_and_enforces_max_depth() {
         let missing_description = execute_tool(
             "Agent",
             &json!({
@@ -2844,6 +3371,22 @@ mod tests {
         )
         .expect_err("blank prompt should fail");
         assert!(missing_prompt.contains("prompt must not be empty"));
+
+        let _guard = env_lock()
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        std::env::set_var("CLAWD_AGENT_DEPTH", "1");
+        let depth_error = execute_tool(
+            "Agent",
+            &json!({
+                "description": "Nested agent",
+                "prompt": "Do nested work.",
+                "max_depth": 1
+            }),
+        )
+        .expect_err("max depth should fail");
+        std::env::remove_var("CLAWD_AGENT_DEPTH");
+        assert!(depth_error.contains("max_depth exceeded"));
     }
 
     #[test]