a3s_code_core/tools/

task.rs

//! Task tools for delegated child runs.
//!
//! The Task tool allows the main agent to delegate specialized work to focused
//! child runs. Each child run gets bounded context and the permissions declared
//! by its agent definition.
//!
//! ## Usage
//!
//! ```json
//! {
//!   "agent": "explore",
//!   "description": "Find authentication code",
//!   "prompt": "Search for files related to user authentication..."
//! }
//! ```

use crate::agent::{AgentConfig, AgentEvent, AgentLoop};
use crate::llm::structured::{generate_blocking, StructuredMode, StructuredRequest};
use crate::llm::LlmClient;
use crate::mcp::manager::McpManager;
use crate::orchestration::{AgentExecutor, AgentStepSpec, StepOutcome};
use crate::subagent::AgentRegistry;
use crate::tools::types::{Tool, ToolContext, ToolOutput};
use anyhow::{Context, Result};
use async_trait::async_trait;
use serde::{Deserialize, Serialize};
use std::path::PathBuf;
use std::sync::Arc;
use tokio::sync::broadcast;

const TASK_OUTPUT_CONTEXT_LIMIT: usize = 4_000;
const TASK_OUTPUT_CONTEXT_HEAD: usize = 3_000;
const TASK_OUTPUT_CONTEXT_TAIL: usize = 800;

/// Task tool parameters
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct TaskParams {
    /// Agent type to use (explore, general, plan, verification, review, etc.)
    pub agent: String,
    /// Short description of the task (for display)
    pub description: String,
    /// Detailed prompt for the agent
    pub prompt: String,
    /// Optional: run in background (default: false)
    #[serde(default)]
    pub background: bool,
    /// Optional: maximum steps for this task
    #[serde(skip_serializing_if = "Option::is_none")]
    pub max_steps: Option<usize>,
}

/// Task tool result
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct TaskResult {
    /// Task output from the delegated child run.
    pub output: String,
    /// Child session ID
    pub session_id: String,
    /// Agent type used
    pub agent: String,
    /// Whether the task succeeded
    pub success: bool,
    /// Task ID for tracking
    pub task_id: String,
}

fn compact_task_output(output: &str) -> (String, bool) {
    if output.len() <= TASK_OUTPUT_CONTEXT_LIMIT {
        return (output.to_string(), false);
    }

    let head = crate::text::truncate_utf8(output, TASK_OUTPUT_CONTEXT_HEAD);
    let tail_start = output
        .char_indices()
        .find_map(|(idx, _)| {
            if output.len().saturating_sub(idx) <= TASK_OUTPUT_CONTEXT_TAIL {
                Some(idx)
            } else {
                None
            }
        })
        .unwrap_or(output.len());
    let tail = &output[tail_start..];

    (
        format!(
            "{}\n\n[{} bytes omitted from delegated task output]\n\n{}",
            head,
            output.len().saturating_sub(head.len() + tail.len()),
            tail
        ),
        true,
    )
}

/// Translate selected child-loop events into a `SubagentProgress` milestone
/// for the parent broadcast. Returns `None` for events that aren't worth
/// surfacing as progress (text deltas, tool starts, subagent events from
/// nested delegation, etc.).
///
/// Currently emits progress for:
/// - `ToolEnd`   → `status = "tool_completed"`,
///   `metadata = { tool, exit_code, output_bytes, error_kind? }`
/// - `TurnEnd`   → `status = "turn_completed"`,
///   `metadata = { turn, total_tokens, prompt_tokens, completion_tokens }`
fn synthesize_subagent_progress(
    event: &AgentEvent,
    task_id: &str,
    session_id: &str,
) -> Option<AgentEvent> {
    match event {
        AgentEvent::ToolEnd {
            name,
            output,
            exit_code,
            error_kind,
            ..
        } => {
            let mut metadata = serde_json::json!({
                "tool": name,
                "exit_code": exit_code,
                "output_bytes": output.len(),
            });
            if let Some(kind) = error_kind {
                metadata["error_kind"] =
                    serde_json::to_value(kind).unwrap_or(serde_json::Value::Null);
            }
            Some(AgentEvent::SubagentProgress {
                task_id: task_id.to_string(),
                session_id: session_id.to_string(),
                status: "tool_completed".to_string(),
                metadata,
            })
        }
        AgentEvent::TurnEnd { turn, usage } => Some(AgentEvent::SubagentProgress {
            task_id: task_id.to_string(),
            session_id: session_id.to_string(),
            status: "turn_completed".to_string(),
            metadata: serde_json::json!({
                "turn": turn,
                "total_tokens": usage.total_tokens,
                "prompt_tokens": usage.prompt_tokens,
                "completion_tokens": usage.completion_tokens,
            }),
        }),
        _ => None,
    }
}

fn task_artifact_id(result: &TaskResult) -> String {
    format!("task-output:{}", result.task_id)
}

fn task_artifact_uri(result: &TaskResult) -> String {
    format!(
        "a3s://tasks/{}/runs/{}/output",
        result.session_id, result.task_id
    )
}

fn epoch_ms() -> u64 {
    use std::time::{SystemTime, UNIX_EPOCH};
    SystemTime::now()
        .duration_since(UNIX_EPOCH)
        .map(|duration| duration.as_millis() as u64)
        .unwrap_or(0)
}

fn format_task_result_for_context(result: &TaskResult) -> (String, bool) {
    let (output, truncated) = compact_task_output(&result.output);
    let status = if result.success {
        "completed"
    } else {
        "failed"
    };
    let artifact_id = task_artifact_id(result);
    let artifact_uri = task_artifact_uri(result);
    let mut formatted = format!(
        "Task {status}: {}\nAgent: {}\nSession: {}\nTask ID: {}\nArtifact ID: {}\nArtifact URI: {}\n",
        result.task_id, result.agent, result.session_id, result.task_id, artifact_id, artifact_uri
    );
    if truncated {
        formatted.push_str(
            "Output excerpt: truncated for parent context. Use the artifact URI or child run session/events if exact omitted content is needed.\n",
        );
    } else {
        formatted.push_str("Output:\n");
    }
    formatted.push_str(&output);
    (formatted, truncated)
}

/// Task executor for delegated child runs.
pub struct TaskExecutor {
    /// Agent registry for looking up agent definitions
    registry: Arc<AgentRegistry>,
    /// LLM client used to power child agent loops
    llm_client: Arc<dyn LlmClient>,
    /// Workspace path shared with child agents
    workspace: String,
    /// Optional MCP manager for registering MCP tools in child sessions
    mcp_manager: Option<Arc<McpManager>>,
    /// Parent capabilities to inherit into child runs.
    parent_context: Option<crate::child_run::ChildRunContext>,
    max_parallel_tasks: usize,
    /// Optional shared tracker — when present each task registers a
    /// `CancellationToken` so callers can cancel by `task_id`.
    subagent_tracker: Option<Arc<crate::subagent_task_tracker::InMemorySubagentTaskTracker>>,
}

impl TaskExecutor {
    /// Create a new task executor
    pub fn new(
        registry: Arc<AgentRegistry>,
        llm_client: Arc<dyn LlmClient>,
        workspace: String,
    ) -> Self {
        Self {
            registry,
            llm_client,
            workspace,
            mcp_manager: None,
            parent_context: None,
            max_parallel_tasks: crate::agent::DEFAULT_MAX_PARALLEL_TASKS,
            subagent_tracker: None,
        }
    }

    /// Create a new task executor with MCP manager for tool inheritance
    pub fn with_mcp(
        registry: Arc<AgentRegistry>,
        llm_client: Arc<dyn LlmClient>,
        workspace: String,
        mcp_manager: Arc<McpManager>,
    ) -> Self {
        Self {
            registry,
            llm_client,
            workspace,
            mcp_manager: Some(mcp_manager),
            parent_context: None,
            max_parallel_tasks: crate::agent::DEFAULT_MAX_PARALLEL_TASKS,
            subagent_tracker: None,
        }
    }

    /// Set parent session capabilities to inherit into child runs.
    pub fn with_parent_context(mut self, ctx: crate::child_run::ChildRunContext) -> Self {
        if let Some(max_parallel_tasks) = ctx.max_parallel_tasks {
            self.max_parallel_tasks = max_parallel_tasks.max(1);
        }
        self.parent_context = Some(ctx);
        self
    }

    pub fn with_max_parallel_tasks(mut self, max_parallel_tasks: usize) -> Self {
        self.max_parallel_tasks = max_parallel_tasks.max(1);
        self
    }

    /// Share a tracker with this executor. When set, each task registers
    /// a `CancellationToken` against the tracker so the parent session
    /// can cancel by `task_id`.
    pub fn with_subagent_tracker(
        mut self,
        tracker: Arc<crate::subagent_task_tracker::InMemorySubagentTaskTracker>,
    ) -> Self {
        self.subagent_tracker = Some(tracker);
        self
    }

    /// Execute a task by spawning an isolated child AgentLoop.
    ///
    /// `parent_session_id` flows into the emitted `SubagentStart`/`SubagentEnd`
    /// events so dashboards can associate child runs with the parent session.
    pub async fn execute(
        &self,
        params: TaskParams,
        event_tx: Option<broadcast::Sender<AgentEvent>>,
        parent_session_id: Option<&str>,
    ) -> Result<TaskResult> {
        let task_id = format!("task-{}", uuid::Uuid::new_v4());
        self.execute_with_task_id(task_id, params, event_tx, parent_session_id, true)
            .await
    }

    /// Execute a task using a caller-supplied task id. Used by `execute_background`
    /// so the synchronously-returned task id matches the one in lifecycle events.
    /// When `emit_start` is `false` the caller is responsible for emitting
    /// `SubagentStart` themselves (e.g. to avoid a race against a tracker query).
    pub async fn execute_with_task_id(
        &self,
        task_id: String,
        params: TaskParams,
        event_tx: Option<broadcast::Sender<AgentEvent>>,
        parent_session_id: Option<&str>,
        emit_start: bool,
    ) -> Result<TaskResult> {
        let session_id = format!("task-run-{}", task_id);
        let started_ms = epoch_ms();

        let agent = self
            .registry
            .get(&params.agent)
            .context(format!("Unknown agent type: '{}'", params.agent))?;

        if emit_start {
            if let Some(ref tx) = event_tx {
                let _ = tx.send(AgentEvent::SubagentStart {
                    task_id: task_id.clone(),
                    session_id: session_id.clone(),
                    parent_session_id: parent_session_id.unwrap_or_default().to_string(),
                    agent: params.agent.clone(),
                    description: params.description.clone(),
                    started_ms,
                });
            }
        }

        // Build a child ToolExecutor. Task tools are intentionally omitted
        // here to prevent unlimited delegation nesting.
        let child_executor = if let Some(ref parent_ctx) = self.parent_context {
            if let Some(ref services) = parent_ctx.workspace_services {
                crate::tools::ToolExecutor::new_with_workspace_services_and_artifact_limits(
                    self.workspace.clone(),
                    Arc::clone(services),
                    crate::tools::ArtifactStoreLimits::default(),
                )
            } else {
                crate::tools::ToolExecutor::new(self.workspace.clone())
            }
        } else {
            crate::tools::ToolExecutor::new(self.workspace.clone())
        };

        // Register MCP tools so child agents can access MCP servers.
        if let Some(ref mcp) = self.mcp_manager {
            let all_tools = mcp.get_all_tools().await;
            let mut by_server: std::collections::HashMap<
                String,
                Vec<crate::mcp::protocol::McpTool>,
            > = std::collections::HashMap::new();
            for (server, tool) in all_tools {
                by_server.entry(server).or_default().push(tool);
            }
            for (server_name, tools) in by_server {
                let wrappers =
                    crate::mcp::tools::create_mcp_tools(&server_name, tools, Arc::clone(mcp));
                for wrapper in wrappers {
                    child_executor.register_dynamic_tool(wrapper);
                }
            }
        }

        let child_executor = Arc::new(child_executor);

        let mut child_config = AgentConfig {
            tools: child_executor.definitions(),
            ..AgentConfig::default()
        };
        agent.apply_to(&mut child_config);
        if let Some(ref parent_ctx) = self.parent_context {
            parent_ctx.apply_to(&mut child_config);
        }
        if let Some(max_steps) = params.max_steps {
            child_config.max_tool_rounds = max_steps;
        }

        let mut tool_context =
            ToolContext::new(PathBuf::from(&self.workspace)).with_session_id(session_id.clone());
        if let Some(ref parent_ctx) = self.parent_context {
            if let Some(ref services) = parent_ctx.workspace_services {
                tool_context = tool_context.with_workspace_services(Arc::clone(services));
            }
        }

        let agent_loop = AgentLoop::new(
            Arc::clone(&self.llm_client),
            child_executor,
            tool_context,
            child_config,
        );

        // Create an mpsc channel for the child agent and forward events to broadcast.
        // Selected child events (ToolEnd, TurnEnd) are also surfaced to the parent
        // broadcast as synthetic `SubagentProgress` events so dashboards can observe
        // mid-task milestones without subscribing to the raw event stream.
        let child_event_tx = if let Some(ref broadcast_tx) = event_tx {
            let (mpsc_tx, mut mpsc_rx) = tokio::sync::mpsc::channel(100);
            let broadcast_tx_clone = broadcast_tx.clone();
            let progress_task_id = task_id.clone();
            let progress_session_id = session_id.clone();

            tokio::spawn(async move {
                while let Some(event) = mpsc_rx.recv().await {
                    if let Some(progress) = synthesize_subagent_progress(
                        &event,
                        &progress_task_id,
                        &progress_session_id,
                    ) {
                        let _ = broadcast_tx_clone.send(progress);
                    }
                    let _ = broadcast_tx_clone.send(event);
                }
            });

            Some(mpsc_tx)
        } else {
            None
        };

        // Register a CancellationToken with the tracker (if shared) so the
        // parent session's `cancel_subagent_task` can interrupt this run.
        let cancel_token = tokio_util::sync::CancellationToken::new();
        if let Some(ref tracker) = self.subagent_tracker {
            tracker
                .register_canceller(&task_id, cancel_token.clone())
                .await;
        }

        let (output, success) = match agent_loop
            .execute_with_session(
                &[],
                &params.prompt,
                Some(&session_id),
                child_event_tx,
                Some(&cancel_token),
            )
            .await
        {
            Ok(result) => (result.text, true),
            Err(e) if cancel_token.is_cancelled() => {
                (format!("Task cancelled by caller: {}", e), false)
            }
            Err(e) => (format!("Task failed: {}", e), false),
        };

        if let Some(ref tracker) = self.subagent_tracker {
            tracker.clear_canceller(&task_id).await;
        }

        if let Some(ref tx) = event_tx {
            let _ = tx.send(AgentEvent::SubagentEnd {
                task_id: task_id.clone(),
                session_id: session_id.clone(),
                agent: params.agent.clone(),
                output: output.clone(),
                success,
                finished_ms: epoch_ms(),
            });
        }

        Ok(TaskResult {
            output,
            session_id,
            agent: params.agent,
            success,
            task_id,
        })
    }

    /// Execute a task in the background.
    ///
    /// Returns immediately with the task ID; the same id is used in the emitted
    /// `SubagentStart`/`SubagentEnd` events so callers can correlate. Pre-emits
    /// `SubagentStart` synchronously when an event channel is available so a
    /// caller that queries the subagent task tracker right after this call
    /// observes the task in `Running` state without a race window.
    pub fn execute_background(
        self: Arc<Self>,
        params: TaskParams,
        event_tx: Option<broadcast::Sender<AgentEvent>>,
        parent_session_id: Option<String>,
    ) -> String {
        let task_id = format!("task-{}", uuid::Uuid::new_v4());
        let session_id = format!("task-run-{}", task_id);

        if let Some(ref tx) = event_tx {
            let _ = tx.send(AgentEvent::SubagentStart {
                task_id: task_id.clone(),
                session_id,
                parent_session_id: parent_session_id.clone().unwrap_or_default(),
                agent: params.agent.clone(),
                description: params.description.clone(),
                started_ms: epoch_ms(),
            });
        }

        let task_id_for_spawn = task_id.clone();
        let task_id_for_log = task_id.clone();
        tokio::spawn(async move {
            if let Err(e) = self
                .execute_with_task_id(
                    task_id_for_spawn,
                    params,
                    event_tx,
                    parent_session_id.as_deref(),
                    false,
                )
                .await
            {
                tracing::error!("Background task {} failed: {}", task_id_for_log, e);
            }
        });

        task_id
    }

    /// Execute multiple tasks in parallel.
    ///
    /// Spawns all tasks concurrently and waits for all to complete.
    /// Returns results in the same order as the input tasks. Routed through
    /// the [`AgentExecutor`](crate::orchestration::AgentExecutor) seam so the
    /// same fan-out works whether steps run locally (default) or are placed
    /// on remote nodes by a host.
    pub async fn execute_parallel(
        self: &Arc<Self>,
        tasks: Vec<TaskParams>,
        event_tx: Option<broadcast::Sender<AgentEvent>>,
        parent_session_id: Option<&str>,
    ) -> Vec<TaskResult> {
        let parent = parent_session_id.map(|s| s.to_string());
        let specs = tasks
            .into_iter()
            .map(|params| AgentStepSpec {
                task_id: format!("task-{}", uuid::Uuid::new_v4()),
                agent: params.agent,
                description: params.description,
                prompt: params.prompt,
                max_steps: params.max_steps,
                parent_session_id: parent.clone(),
                output_schema: None,
            })
            .collect();

        let executor: Arc<dyn AgentExecutor> = Arc::<Self>::clone(self);
        crate::orchestration::execute_steps_parallel(executor, specs, event_tx)
            .await
            .into_iter()
            .map(TaskResult::from)
            .collect()
    }
}

impl From<TaskResult> for StepOutcome {
    fn from(r: TaskResult) -> Self {
        StepOutcome {
            task_id: r.task_id,
            session_id: r.session_id,
            agent: r.agent,
            output: r.output,
            success: r.success,
            structured: None,
        }
    }
}

impl From<StepOutcome> for TaskResult {
    fn from(o: StepOutcome) -> Self {
        TaskResult {
            output: o.output,
            session_id: o.session_id,
            agent: o.agent,
            success: o.success,
            task_id: o.task_id,
        }
    }
}

/// The local, in-process executor: every step runs as a child `AgentLoop` on
/// this node's tokio runtime. This is the default; a host substitutes
/// its own [`AgentExecutor`] to place steps across a cluster.
#[async_trait]
impl AgentExecutor for TaskExecutor {
    async fn execute_step(
        &self,
        spec: AgentStepSpec,
        event_tx: Option<broadcast::Sender<AgentEvent>>,
    ) -> StepOutcome {
        let agent = spec.agent.clone();
        let task_id = spec.task_id.clone();
        let output_schema = spec.output_schema.clone();
        let params = TaskParams {
            agent: spec.agent,
            description: spec.description,
            prompt: spec.prompt,
            background: false,
            max_steps: spec.max_steps,
        };
        let mut outcome: StepOutcome = match self
            .execute_with_task_id(
                task_id.clone(),
                params,
                event_tx,
                spec.parent_session_id.as_deref(),
                true,
            )
            .await
        {
            Ok(result) => result.into(),
            Err(e) => return StepOutcome::failed(task_id, agent, format!("Task failed: {e}")),
        };

        // When the step requested structured output, coerce the (succeeded)
        // free-text result to the schema. A coercion failure demotes the step
        // to unsuccessful so callers never treat unvalidated text as the
        // promised object.
        if outcome.success {
            if let Some(schema) = output_schema {
                match self.coerce_to_schema(&outcome.output, schema).await {
                    Ok(object) => outcome.structured = Some(object),
                    Err(e) => {
                        outcome.success = false;
                        outcome.output =
                            format!("{}\n\n[structured output failed: {e}]", outcome.output);
                    }
                }
            }
        }
        outcome
    }

    fn concurrency_hint(&self) -> usize {
        self.max_parallel_tasks
    }
}

impl TaskExecutor {
    /// Coerce a step's free-text output into a JSON object validated against
    /// `schema`, reusing the structured-output machinery (Tool mode — the most
    /// portable across providers, with built-in repair). This is one extra LLM
    /// call beyond the step's own run.
    async fn coerce_to_schema(
        &self,
        output: &str,
        schema: serde_json::Value,
    ) -> Result<serde_json::Value> {
        let req = StructuredRequest {
            prompt: format!(
                "Convert the following task result into a single JSON object that conforms to \
                 the required schema. Use only information present in the result.\n\n\
                 --- TASK RESULT ---\n{output}"
            ),
            system: Some(
                "You output exactly one JSON object matching the provided schema.".to_string(),
            ),
            schema,
            schema_name: "step_output".to_string(),
            schema_description: None,
            // Tool mode works on every provider that supports tool use and
            // does not depend on response_format wiring.
            mode: StructuredMode::Tool,
            max_repair_attempts: 2,
        };
        let result = generate_blocking(&*self.llm_client, &req).await?;
        Ok(result.object)
    }
}

/// Get the JSON schema for TaskParams
pub fn task_params_schema() -> serde_json::Value {
    serde_json::json!({
        "type": "object",
        "additionalProperties": false,
        "properties": {
            "agent": {
                "type": "string",
                "description": "Required. Canonical agent type to use (for example: explore, general, plan, verification, review). Always provide this exact field name: 'agent'."
            },
            "description": {
                "type": "string",
                "description": "Required. Short task label for display and tracking. Always provide this exact field name: 'description'."
            },
            "prompt": {
                "type": "string",
                "description": "Required. Detailed instruction for the delegated child run. Always provide this exact field name: 'prompt'."
            },
            "background": {
                "type": "boolean",
                "description": "Optional. Run the task in the background. Default: false.",
                "default": false
            },
            "max_steps": {
                "type": "integer",
                "description": "Optional. Maximum number of steps for this task."
            }
        },
        "required": ["agent", "description", "prompt"],
        "examples": [
            {
                "agent": "explore",
                "description": "Find Rust files",
                "prompt": "Search the workspace for Rust files and summarize the layout."
            },
            {
                "agent": "general",
                "description": "Investigate test failure",
                "prompt": "Inspect the failing tests and explain the root cause.",
                "max_steps": 6
            }
        ]
    })
}

/// TaskTool wraps TaskExecutor as a Tool for registration in ToolExecutor.
/// This allows the LLM to delegate tasks through the standard tool interface.
pub struct TaskTool {
    executor: Arc<TaskExecutor>,
}

impl TaskTool {
    /// Create a new TaskTool
    pub fn new(executor: Arc<TaskExecutor>) -> Self {
        Self { executor }
    }
}

#[async_trait]
impl Tool for TaskTool {
    fn name(&self) -> &str {
        "task"
    }

    fn description(&self) -> &str {
        "Delegate a bounded task to a specialized child run. Built-in agents: explore (read-only codebase and web evidence search), general/general-purpose (full access multi-step), plan (read-only planning), verification (adversarial validation), review (code review). Custom agents from agent_dirs and .a3s/agents are also available; .claude/agents is read for compatibility."
    }

    fn parameters(&self) -> serde_json::Value {
        task_params_schema()
    }

    async fn execute(&self, args: &serde_json::Value, ctx: &ToolContext) -> Result<ToolOutput> {
        let params: TaskParams =
            serde_json::from_value(args.clone()).context("Invalid task parameters")?;

        if params.background {
            let task_id = Arc::clone(&self.executor).execute_background(
                params,
                ctx.agent_event_tx.clone(),
                ctx.session_id.clone(),
            );
            return Ok(ToolOutput::success(format!(
                "Task started in background. Task ID: {}",
                task_id
            )));
        }

        let result = self
            .executor
            .execute(
                params,
                ctx.agent_event_tx.clone(),
                ctx.session_id.as_deref(),
            )
            .await?;
        let (content, truncated) = format_task_result_for_context(&result);
        let metadata = serde_json::json!({
            "task_id": result.task_id,
            "session_id": result.session_id,
            "agent": result.agent,
            "success": result.success,
            "output_bytes": result.output.len(),
            "truncated_for_context": truncated,
            "artifact_id": task_artifact_id(&result),
            "artifact_uri": task_artifact_uri(&result),
        });

        if result.success {
            Ok(ToolOutput::success(content).with_metadata(metadata))
        } else {
            Ok(ToolOutput::error(content).with_metadata(metadata))
        }
    }
}

/// Parameters for parallel task execution
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct ParallelTaskParams {
    /// List of tasks to execute concurrently
    pub tasks: Vec<TaskParams>,
}

/// Get the JSON schema for ParallelTaskParams
pub fn parallel_task_params_schema() -> serde_json::Value {
    serde_json::json!({
        "type": "object",
        "additionalProperties": false,
        "properties": {
            "tasks": {
                "type": "array",
                "description": "List of tasks to execute in parallel. Each task runs as an independent delegated child run concurrently.",
                "items": {
                    "type": "object",
                    "additionalProperties": false,
                    "properties": {
                        "agent": {
                            "type": "string",
                            "description": "Required. Canonical agent type for this task."
                        },
                        "description": {
                            "type": "string",
                            "description": "Required. Short task label for display and tracking."
                        },
                        "prompt": {
                            "type": "string",
                            "description": "Required. Detailed instruction for the delegated child run."
                        }
                    },
                    "required": ["agent", "description", "prompt"]
                },
                "minItems": 1
            }
        },
        "required": ["tasks"],
        "examples": [
            {
                "tasks": [
                    {
                        "agent": "explore",
                        "description": "Find Rust files",
                        "prompt": "List Rust files under src/."
                    },
                    {
                        "agent": "explore",
                        "description": "Find tests",
                        "prompt": "List test files and summarize their purpose."
                    }
                ]
            }
        ]
    })
}

/// ParallelTaskTool allows the LLM to fan out multiple delegated tasks concurrently.
///
/// All tasks execute in parallel and the tool returns when all complete.
pub struct ParallelTaskTool {
    executor: Arc<TaskExecutor>,
}

impl ParallelTaskTool {
    /// Create a new ParallelTaskTool
    pub fn new(executor: Arc<TaskExecutor>) -> Self {
        Self { executor }
    }
}

#[async_trait]
impl Tool for ParallelTaskTool {
    fn name(&self) -> &str {
        "parallel_task"
    }

    fn description(&self) -> &str {
        "Fan out 2 or more INDEPENDENT subtasks as delegated child runs that execute concurrently; results are returned when all complete. Use this only when the work genuinely splits into branches that can be investigated or implemented separately (e.g. inspect several unrelated modules at once, or run review and verification in parallel). Do NOT use it for trivial, conversational, or single-step requests, or for steps that depend on one another — handle those directly. Built-in agents: explore (read-only codebase and web evidence search), general/general-purpose (full access multi-step), plan (read-only planning), verification (adversarial validation), review (code review). Custom agents from agent_dirs and .a3s/agents are also available; .claude/agents is read for compatibility."
    }

    fn parameters(&self) -> serde_json::Value {
        parallel_task_params_schema()
    }

    async fn execute(&self, args: &serde_json::Value, ctx: &ToolContext) -> Result<ToolOutput> {
        let params: ParallelTaskParams =
            serde_json::from_value(args.clone()).context("Invalid parallel task parameters")?;

        if params.tasks.is_empty() {
            return Ok(ToolOutput::error("No tasks provided".to_string()));
        }

        let task_count = params.tasks.len();

        let results = self
            .executor
            .execute_parallel(
                params.tasks,
                ctx.agent_event_tx.clone(),
                ctx.session_id.as_deref(),
            )
            .await;

        // Format results with compact per-task excerpts for parent context.
        let mut output = format!("Executed {} tasks in parallel:\n\n", task_count);
        let mut metadata_results = Vec::new();
        for (i, result) in results.iter().enumerate() {
            let status = if result.success { "[OK]" } else { "[ERR]" };
            let (formatted, truncated) = format_task_result_for_context(result);
            metadata_results.push(serde_json::json!({
                "task_id": result.task_id,
                "session_id": result.session_id,
                "agent": result.agent,
                "success": result.success,
                "output": formatted.clone(),
                "output_bytes": result.output.len(),
                "truncated_for_context": truncated,
                "artifact_id": task_artifact_id(result),
                "artifact_uri": task_artifact_uri(result),
            }));
            output.push_str(&format!(
                "--- Task {} ({}) {} ---\n{}\n\n",
                i + 1,
                result.agent,
                status,
                formatted
            ));
        }

        let all_success = results.iter().all(|result| result.success);
        let output = if all_success {
            ToolOutput::success(output)
        } else {
            ToolOutput::error(output)
        };

        Ok(output.with_metadata(serde_json::json!({
            "task_count": task_count,
            "results": metadata_results,
        })))
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_task_params_deserialize() {
        let json = r#"{
            "agent": "explore",
            "description": "Find auth code",
            "prompt": "Search for authentication files"
        }"#;

        let params: TaskParams = serde_json::from_str(json).unwrap();
        assert_eq!(params.agent, "explore");
        assert_eq!(params.description, "Find auth code");
        assert!(!params.background);
    }

    #[test]
    fn test_task_params_with_background() {
        let json = r#"{
            "agent": "general",
            "description": "Long task",
            "prompt": "Do something complex",
            "background": true
        }"#;

        let params: TaskParams = serde_json::from_str(json).unwrap();
        assert!(params.background);
    }

    #[test]
    fn test_task_params_with_max_steps() {
        let json = r#"{
            "agent": "plan",
            "description": "Planning task",
            "prompt": "Create a plan",
            "max_steps": 10
        }"#;

        let params: TaskParams = serde_json::from_str(json).unwrap();
        assert_eq!(params.agent, "plan");
        assert_eq!(params.max_steps, Some(10));
        assert!(!params.background);
    }

    #[test]
    fn test_task_params_all_fields() {
        let json = r#"{
            "agent": "general",
            "description": "Complex task",
            "prompt": "Do everything",
            "background": true,
            "max_steps": 20
        }"#;

        let params: TaskParams = serde_json::from_str(json).unwrap();
        assert_eq!(params.agent, "general");
        assert_eq!(params.description, "Complex task");
        assert_eq!(params.prompt, "Do everything");
        assert!(params.background);
        assert_eq!(params.max_steps, Some(20));
    }

    #[test]
    fn test_task_params_missing_required_field() {
        let json = r#"{
            "agent": "explore",
            "description": "Missing prompt"
        }"#;

        let result: Result<TaskParams, _> = serde_json::from_str(json);
        assert!(result.is_err());
    }

    #[test]
    fn test_task_params_serialize() {
        let params = TaskParams {
            agent: "explore".to_string(),
            description: "Test task".to_string(),
            prompt: "Test prompt".to_string(),
            background: false,
            max_steps: Some(5),
        };

        let json = serde_json::to_string(&params).unwrap();
        assert!(json.contains("explore"));
        assert!(json.contains("Test task"));
        assert!(json.contains("Test prompt"));
    }

    #[test]
    fn test_task_params_clone() {
        let params = TaskParams {
            agent: "explore".to_string(),
            description: "Test".to_string(),
            prompt: "Prompt".to_string(),
            background: true,
            max_steps: None,
        };

        let cloned = params.clone();
        assert_eq!(params.agent, cloned.agent);
        assert_eq!(params.description, cloned.description);
        assert_eq!(params.background, cloned.background);
    }

    #[test]
    fn test_task_result_serialize() {
        let result = TaskResult {
            output: "Found 5 files".to_string(),
            session_id: "session-123".to_string(),
            agent: "explore".to_string(),
            success: true,
            task_id: "task-456".to_string(),
        };

        let json = serde_json::to_string(&result).unwrap();
        assert!(json.contains("Found 5 files"));
        assert!(json.contains("explore"));
    }

    #[test]
    fn test_task_result_deserialize() {
        let json = r#"{
            "output": "Task completed",
            "session_id": "sess-789",
            "agent": "general",
            "success": false,
            "task_id": "task-123"
        }"#;

        let result: TaskResult = serde_json::from_str(json).unwrap();
        assert_eq!(result.output, "Task completed");
        assert_eq!(result.session_id, "sess-789");
        assert_eq!(result.agent, "general");
        assert!(!result.success);
        assert_eq!(result.task_id, "task-123");
    }

    #[test]
    fn test_task_result_clone() {
        let result = TaskResult {
            output: "Output".to_string(),
            session_id: "session-1".to_string(),
            agent: "explore".to_string(),
            success: true,
            task_id: "task-1".to_string(),
        };

        let cloned = result.clone();
        assert_eq!(result.output, cloned.output);
        assert_eq!(result.success, cloned.success);
    }

    #[test]
    fn test_compact_task_output_preserves_small_output() {
        let (output, truncated) = compact_task_output("short result");
        assert_eq!(output, "short result");
        assert!(!truncated);
    }

    #[test]
    fn test_format_task_result_for_context_truncates_large_output() {
        let result = TaskResult {
            output: format!("{}TAIL", "x".repeat(TASK_OUTPUT_CONTEXT_LIMIT + 500)),
            session_id: "session-1".to_string(),
            agent: "explore".to_string(),
            success: true,
            task_id: "task-1".to_string(),
        };

        let (formatted, truncated) = format_task_result_for_context(&result);
        assert!(truncated);
        assert!(formatted.contains("Output excerpt"));
        assert!(formatted.contains("bytes omitted"));
        assert!(formatted.contains("Artifact ID: task-output:task-1"));
        assert!(formatted.contains("Artifact URI: a3s://tasks/session-1/runs/task-1/output"));
        assert!(formatted.contains("TAIL"));
        assert!(formatted.len() < result.output.len());
    }

    #[test]
    fn test_task_artifact_reference_is_stable() {
        let result = TaskResult {
            output: "done".to_string(),
            session_id: "session-1".to_string(),
            agent: "explore".to_string(),
            success: true,
            task_id: "task-1".to_string(),
        };

        assert_eq!(task_artifact_id(&result), "task-output:task-1");
        assert_eq!(
            task_artifact_uri(&result),
            "a3s://tasks/session-1/runs/task-1/output"
        );

        let (formatted, truncated) = format_task_result_for_context(&result);
        assert!(!truncated);
        assert!(formatted.contains("Artifact URI: a3s://tasks/session-1/runs/task-1/output"));
    }

    #[test]
    fn test_task_params_schema() {
        let schema = task_params_schema();
        assert_eq!(schema["type"], "object");
        assert_eq!(schema["additionalProperties"], false);
        assert!(schema["properties"]["agent"].is_object());
        assert!(schema["properties"]["prompt"].is_object());
    }

    #[test]
    fn test_task_params_schema_required_fields() {
        let schema = task_params_schema();
        let required = schema["required"].as_array().unwrap();
        assert!(required.contains(&serde_json::json!("agent")));
        assert!(required.contains(&serde_json::json!("description")));
        assert!(required.contains(&serde_json::json!("prompt")));
    }

    #[test]
    fn test_task_params_schema_properties() {
        let schema = task_params_schema();
        let props = &schema["properties"];

        assert_eq!(props["agent"]["type"], "string");
        assert_eq!(props["description"]["type"], "string");
        assert_eq!(props["prompt"]["type"], "string");
        assert_eq!(props["background"]["type"], "boolean");
        assert_eq!(props["background"]["default"], false);
        assert_eq!(props["max_steps"]["type"], "integer");
    }

    #[test]
    fn test_task_params_schema_descriptions() {
        let schema = task_params_schema();
        let props = &schema["properties"];

        assert!(props["agent"]["description"].is_string());
        assert!(props["description"]["description"].is_string());
        assert!(props["prompt"]["description"].is_string());
        assert!(props["background"]["description"].is_string());
        assert!(props["max_steps"]["description"].is_string());
    }

    #[test]
    fn test_task_params_default_background() {
        let params = TaskParams {
            agent: "explore".to_string(),
            description: "Test".to_string(),
            prompt: "Test prompt".to_string(),
            background: false,
            max_steps: None,
        };
        assert!(!params.background);
    }

    #[test]
    fn test_task_params_serialize_skip_none() {
        let params = TaskParams {
            agent: "explore".to_string(),
            description: "Test".to_string(),
            prompt: "Test prompt".to_string(),
            background: false,
            max_steps: None,
        };
        let json = serde_json::to_string(&params).unwrap();
        // max_steps should not appear when None
        assert!(!json.contains("max_steps"));
    }

    #[test]
    fn test_task_params_serialize_with_max_steps() {
        let params = TaskParams {
            agent: "explore".to_string(),
            description: "Test".to_string(),
            prompt: "Test prompt".to_string(),
            background: false,
            max_steps: Some(15),
        };
        let json = serde_json::to_string(&params).unwrap();
        assert!(json.contains("max_steps"));
        assert!(json.contains("15"));
    }

    #[test]
    fn test_task_result_success_true() {
        let result = TaskResult {
            output: "Success".to_string(),
            session_id: "sess-1".to_string(),
            agent: "explore".to_string(),
            success: true,
            task_id: "task-1".to_string(),
        };
        assert!(result.success);
    }

    #[test]
    fn test_task_result_success_false() {
        let result = TaskResult {
            output: "Failed".to_string(),
            session_id: "sess-1".to_string(),
            agent: "explore".to_string(),
            success: false,
            task_id: "task-1".to_string(),
        };
        assert!(!result.success);
    }

    #[test]
    fn test_task_params_empty_strings() {
        let params = TaskParams {
            agent: "".to_string(),
            description: "".to_string(),
            prompt: "".to_string(),
            background: false,
            max_steps: None,
        };
        let json = serde_json::to_string(&params).unwrap();
        let deserialized: TaskParams = serde_json::from_str(&json).unwrap();
        assert_eq!(deserialized.agent, "");
        assert_eq!(deserialized.description, "");
        assert_eq!(deserialized.prompt, "");
    }

    #[test]
    fn test_task_result_empty_output() {
        let result = TaskResult {
            output: "".to_string(),
            session_id: "sess-1".to_string(),
            agent: "explore".to_string(),
            success: true,
            task_id: "task-1".to_string(),
        };
        assert_eq!(result.output, "");
    }

    #[test]
    fn test_task_params_debug_format() {
        let params = TaskParams {
            agent: "explore".to_string(),
            description: "Test".to_string(),
            prompt: "Test prompt".to_string(),
            background: false,
            max_steps: None,
        };
        let debug_str = format!("{:?}", params);
        assert!(debug_str.contains("explore"));
        assert!(debug_str.contains("Test"));
    }

    #[test]
    fn test_task_result_debug_format() {
        let result = TaskResult {
            output: "Output".to_string(),
            session_id: "sess-1".to_string(),
            agent: "explore".to_string(),
            success: true,
            task_id: "task-1".to_string(),
        };
        let debug_str = format!("{:?}", result);
        assert!(debug_str.contains("Output"));
        assert!(debug_str.contains("explore"));
    }

    #[test]
    fn test_task_params_roundtrip() {
        let original = TaskParams {
            agent: "general".to_string(),
            description: "Roundtrip test".to_string(),
            prompt: "Test roundtrip serialization".to_string(),
            background: true,
            max_steps: Some(42),
        };
        let json = serde_json::to_string(&original).unwrap();
        let deserialized: TaskParams = serde_json::from_str(&json).unwrap();
        assert_eq!(original.agent, deserialized.agent);
        assert_eq!(original.description, deserialized.description);
        assert_eq!(original.prompt, deserialized.prompt);
        assert_eq!(original.background, deserialized.background);
        assert_eq!(original.max_steps, deserialized.max_steps);
    }

    #[test]
    fn test_task_result_roundtrip() {
        let original = TaskResult {
            output: "Roundtrip output".to_string(),
            session_id: "sess-roundtrip".to_string(),
            agent: "plan".to_string(),
            success: false,
            task_id: "task-roundtrip".to_string(),
        };
        let json = serde_json::to_string(&original).unwrap();
        let deserialized: TaskResult = serde_json::from_str(&json).unwrap();
        assert_eq!(original.output, deserialized.output);
        assert_eq!(original.session_id, deserialized.session_id);
        assert_eq!(original.agent, deserialized.agent);
        assert_eq!(original.success, deserialized.success);
        assert_eq!(original.task_id, deserialized.task_id);
    }

    #[test]
    fn test_parallel_task_params_deserialize() {
        let json = r#"{
            "tasks": [
                { "agent": "explore", "description": "Find auth", "prompt": "Search auth files" },
                { "agent": "general", "description": "Fix bug", "prompt": "Fix the login bug" }
            ]
        }"#;

        let params: ParallelTaskParams = serde_json::from_str(json).unwrap();
        assert_eq!(params.tasks.len(), 2);
        assert_eq!(params.tasks[0].agent, "explore");
        assert_eq!(params.tasks[1].agent, "general");
    }

    #[test]
    fn test_parallel_task_params_single_task() {
        let json = r#"{
            "tasks": [
                { "agent": "plan", "description": "Plan work", "prompt": "Create a plan" }
            ]
        }"#;

        let params: ParallelTaskParams = serde_json::from_str(json).unwrap();
        assert_eq!(params.tasks.len(), 1);
    }

    #[test]
    fn test_parallel_task_params_empty_tasks() {
        let json = r#"{ "tasks": [] }"#;
        let params: ParallelTaskParams = serde_json::from_str(json).unwrap();
        assert!(params.tasks.is_empty());
    }

    #[test]
    fn test_parallel_task_params_missing_tasks() {
        let json = r#"{}"#;
        let result: Result<ParallelTaskParams, _> = serde_json::from_str(json);
        assert!(result.is_err());
    }

    #[test]
    fn test_parallel_task_params_serialize() {
        let params = ParallelTaskParams {
            tasks: vec![
                TaskParams {
                    agent: "explore".to_string(),
                    description: "Task 1".to_string(),
                    prompt: "Prompt 1".to_string(),
                    background: false,
                    max_steps: None,
                },
                TaskParams {
                    agent: "general".to_string(),
                    description: "Task 2".to_string(),
                    prompt: "Prompt 2".to_string(),
                    background: false,
                    max_steps: Some(10),
                },
            ],
        };
        let json = serde_json::to_string(&params).unwrap();
        assert!(json.contains("explore"));
        assert!(json.contains("general"));
        assert!(json.contains("Prompt 1"));
        assert!(json.contains("Prompt 2"));
    }

    #[test]
    fn test_parallel_task_params_roundtrip() {
        let original = ParallelTaskParams {
            tasks: vec![
                TaskParams {
                    agent: "explore".to_string(),
                    description: "Explore".to_string(),
                    prompt: "Find files".to_string(),
                    background: false,
                    max_steps: None,
                },
                TaskParams {
                    agent: "plan".to_string(),
                    description: "Plan".to_string(),
                    prompt: "Make plan".to_string(),
                    background: false,
                    max_steps: Some(5),
                },
            ],
        };
        let json = serde_json::to_string(&original).unwrap();
        let deserialized: ParallelTaskParams = serde_json::from_str(&json).unwrap();
        assert_eq!(original.tasks.len(), deserialized.tasks.len());
        assert_eq!(original.tasks[0].agent, deserialized.tasks[0].agent);
        assert_eq!(original.tasks[1].agent, deserialized.tasks[1].agent);
        assert_eq!(original.tasks[1].max_steps, deserialized.tasks[1].max_steps);
    }

    #[test]
    fn test_parallel_task_params_clone() {
        let params = ParallelTaskParams {
            tasks: vec![TaskParams {
                agent: "explore".to_string(),
                description: "Test".to_string(),
                prompt: "Prompt".to_string(),
                background: false,
                max_steps: None,
            }],
        };
        let cloned = params.clone();
        assert_eq!(params.tasks.len(), cloned.tasks.len());
        assert_eq!(params.tasks[0].agent, cloned.tasks[0].agent);
    }

    #[test]
    fn test_parallel_task_params_schema() {
        let schema = parallel_task_params_schema();
        assert_eq!(schema["type"], "object");
        assert_eq!(schema["additionalProperties"], false);
        assert!(schema["properties"]["tasks"].is_object());
        assert_eq!(schema["properties"]["tasks"]["type"], "array");
        assert_eq!(schema["properties"]["tasks"]["minItems"], 1);
    }

    #[test]
    fn test_parallel_task_params_schema_required() {
        let schema = parallel_task_params_schema();
        let required = schema["required"].as_array().unwrap();
        assert!(required.contains(&serde_json::json!("tasks")));
    }

    #[test]
    fn test_parallel_task_params_schema_items() {
        let schema = parallel_task_params_schema();
        let items = &schema["properties"]["tasks"]["items"];
        assert_eq!(items["type"], "object");
        assert_eq!(items["additionalProperties"], false);
        let item_required = items["required"].as_array().unwrap();
        assert!(item_required.contains(&serde_json::json!("agent")));
        assert!(item_required.contains(&serde_json::json!("description")));
        assert!(item_required.contains(&serde_json::json!("prompt")));
    }

    #[test]
    fn test_task_schema_examples_use_delegation_core() {
        let task = task_params_schema();
        let task_examples = task["examples"].as_array().unwrap();
        assert_eq!(task_examples[0]["agent"], "explore");
        assert!(task_examples[0].get("task").is_none());

        let parallel = parallel_task_params_schema();
        let parallel_examples = parallel["examples"].as_array().unwrap();
        assert!(!parallel_examples[0]["tasks"].as_array().unwrap().is_empty());
    }

    #[test]
    fn test_parallel_task_params_debug() {
        let params = ParallelTaskParams {
            tasks: vec![TaskParams {
                agent: "explore".to_string(),
                description: "Debug test".to_string(),
                prompt: "Test".to_string(),
                background: false,
                max_steps: None,
            }],
        };
        let debug_str = format!("{:?}", params);
        assert!(debug_str.contains("explore"));
        assert!(debug_str.contains("Debug test"));
    }

    #[test]
    fn test_parallel_task_params_large_count() {
        // Validate that ParallelTaskParams can hold 150 tasks without truncation
        let tasks: Vec<TaskParams> = (0..150)
            .map(|i| TaskParams {
                agent: "explore".to_string(),
                description: format!("Task {}", i),
                prompt: format!("Prompt for task {}", i),
                background: false,
                max_steps: Some(10),
            })
            .collect();

        let params = ParallelTaskParams { tasks };
        let json = serde_json::to_string(&params).unwrap();
        let deserialized: ParallelTaskParams = serde_json::from_str(&json).unwrap();
        assert_eq!(deserialized.tasks.len(), 150);
        assert_eq!(deserialized.tasks[0].description, "Task 0");
        assert_eq!(deserialized.tasks[149].description, "Task 149");
    }

    #[test]
    fn test_task_params_max_steps_zero() {
        // max_steps = 0 is a valid edge case (callers decide enforcement)
        let params = TaskParams {
            agent: "explore".to_string(),
            description: "Edge case".to_string(),
            prompt: "Zero steps".to_string(),
            background: false,
            max_steps: Some(0),
        };
        let json = serde_json::to_string(&params).unwrap();
        let deserialized: TaskParams = serde_json::from_str(&json).unwrap();
        assert_eq!(deserialized.max_steps, Some(0));
    }

    #[test]
    fn test_parallel_task_params_all_background() {
        let tasks: Vec<TaskParams> = (0..5)
            .map(|i| TaskParams {
                agent: "general".to_string(),
                description: format!("BG task {}", i),
                prompt: "Run in background".to_string(),
                background: true,
                max_steps: None,
            })
            .collect();
        let params = ParallelTaskParams { tasks };
        for task in &params.tasks {
            assert!(task.background);
        }
    }

    #[test]
    fn test_task_params_rejects_permissive_field() {
        let json = r#"{
            "agent": "general",
            "description": "Legacy field rejection",
            "prompt": "Verify legacy fields are rejected",
            "permissive": true
        }"#;

        let result: Result<TaskParams, _> = serde_json::from_str(json);
        assert!(result.is_err());
    }

    #[test]
    fn test_task_params_schema_hides_permissive_field() {
        let schema = task_params_schema();
        let props = &schema["properties"];

        assert!(props.get("permissive").is_none());
    }

    // ========================================================================
    // Contract tests — verify task delegation with MockLlmClient (no network)
    // ========================================================================

    use crate::agent::tests::MockLlmClient;
    use crate::llm::{ContentBlock, LlmResponse, Message, StreamEvent, TokenUsage, ToolDefinition};
    use crate::permissions::PermissionPolicy;
    use crate::subagent::AgentRegistry;
    use std::sync::atomic::{AtomicUsize, Ordering};
    use std::time::Duration;
    use tokio::sync::{mpsc, Barrier};

    fn text_response(text: impl Into<String>) -> LlmResponse {
        LlmResponse {
            message: Message {
                role: "assistant".to_string(),
                content: vec![ContentBlock::Text { text: text.into() }],
                reasoning_content: None,
            },
            usage: TokenUsage {
                prompt_tokens: 10,
                completion_tokens: 5,
                total_tokens: 15,
                cache_read_tokens: None,
                cache_write_tokens: None,
            },
            stop_reason: Some("end_turn".to_string()),
            token_logprobs: Vec::new(),
            meta: None,
        }
    }

    fn pre_analysis_response(messages: &[Message]) -> LlmResponse {
        let prompt = last_text(messages);
        let response = serde_json::json!({
            "intent": "GeneralPurpose",
            "requires_planning": false,
            "goal": {
                "description": prompt,
                "success_criteria": []
            },
            "execution_plan": {
                "complexity": "Simple",
                "steps": [{
                    "id": "step-1",
                    "description": prompt,
                    "tool": null,
                    "dependencies": [],
                    "success_criteria": "Complete the request"
                }],
                "required_tools": []
            },
            "optimized_input": prompt
        });
        text_response(response.to_string())
    }

    fn last_text(messages: &[Message]) -> String {
        messages
            .last()
            .and_then(|message| {
                message.content.iter().find_map(|block| {
                    if let ContentBlock::Text { text } = block {
                        Some(text.clone())
                    } else {
                        None
                    }
                })
            })
            .unwrap_or_default()
    }

    /// Client for the schema-coercion tests. The agent's own turn returns
    /// plain text (which ends the loop); the structured-output coercion call
    /// — recognizable by the injected `step_output` tool — returns a tool call
    /// carrying the object.
    struct SchemaCoercionClient;

    #[async_trait::async_trait]
    impl LlmClient for SchemaCoercionClient {
        async fn complete(
            &self,
            messages: &[Message],
            system: Option<&str>,
            tools: &[ToolDefinition],
        ) -> Result<LlmResponse> {
            if system == Some(crate::prompts::PRE_ANALYSIS_SYSTEM) {
                return Ok(pre_analysis_response(messages));
            }
            // The structured-output coercion injects a synthetic tool named
            // `emit_<schema_name>` (here `emit_step_output`).
            if tools.iter().any(|t| t.name == "emit_step_output") {
                return Ok(MockLlmClient::tool_call_response(
                    "coerce-1",
                    "emit_step_output",
                    serde_json::json!({ "verdict": "ok" }),
                ));
            }
            Ok(text_response("The verdict is ok."))
        }

        async fn complete_streaming(
            &self,
            _messages: &[Message],
            _system: Option<&str>,
            _tools: &[ToolDefinition],
            _cancel_token: tokio_util::sync::CancellationToken,
        ) -> Result<mpsc::Receiver<StreamEvent>> {
            anyhow::bail!("streaming is not used by schema coercion tests")
        }
    }

    fn verdict_schema() -> serde_json::Value {
        serde_json::json!({
            "type": "object",
            "properties": { "verdict": { "type": "string" } },
            "required": ["verdict"]
        })
    }

    #[tokio::test]
    async fn execute_step_with_schema_coerces_structured_output() {
        let workspace = tempfile::tempdir().unwrap();
        let executor = TaskExecutor::new(
            Arc::new(AgentRegistry::new()),
            Arc::new(SchemaCoercionClient),
            workspace.path().to_string_lossy().to_string(),
        );
        let spec = AgentStepSpec::new("step-1", "general", "assess", "Assess the thing.")
            .with_output_schema(verdict_schema());

        let outcome = executor.execute_step(spec, None).await;

        assert!(outcome.success, "step should succeed: {}", outcome.output);
        assert_eq!(
            outcome.structured,
            Some(serde_json::json!({ "verdict": "ok" })),
            "a schema'd step returns the validated object in `structured`"
        );
    }

    #[tokio::test]
    async fn execute_step_without_schema_has_no_structured_output() {
        let workspace = tempfile::tempdir().unwrap();
        let executor = TaskExecutor::new(
            Arc::new(AgentRegistry::new()),
            Arc::new(SchemaCoercionClient),
            workspace.path().to_string_lossy().to_string(),
        );
        let spec = AgentStepSpec::new("step-2", "general", "assess", "Assess the thing.");

        let outcome = executor.execute_step(spec, None).await;

        assert!(outcome.success, "step should succeed: {}", outcome.output);
        assert_eq!(
            outcome.structured, None,
            "no schema requested → no structured output, no coercion call"
        );
    }

    /// The agent's turn returns text; the coercion call (`emit_step_output`)
    /// always returns an object that VIOLATES the schema, so `generate_blocking`
    /// exhausts its repairs and bails.
    struct SchemaFailClient;

    #[async_trait::async_trait]
    impl LlmClient for SchemaFailClient {
        async fn complete(
            &self,
            messages: &[Message],
            system: Option<&str>,
            tools: &[ToolDefinition],
        ) -> Result<LlmResponse> {
            if system == Some(crate::prompts::PRE_ANALYSIS_SYSTEM) {
                return Ok(pre_analysis_response(messages));
            }
            if tools.iter().any(|t| t.name == "emit_step_output") {
                // `{}` is missing the required `verdict` field → schema invalid.
                return Ok(MockLlmClient::tool_call_response(
                    "coerce-fail",
                    "emit_step_output",
                    serde_json::json!({}),
                ));
            }
            Ok(text_response("some answer"))
        }

        async fn complete_streaming(
            &self,
            _messages: &[Message],
            _system: Option<&str>,
            _tools: &[ToolDefinition],
            _cancel_token: tokio_util::sync::CancellationToken,
        ) -> Result<mpsc::Receiver<StreamEvent>> {
            anyhow::bail!("streaming unused")
        }
    }

    #[tokio::test]
    async fn execute_step_with_schema_demotes_step_on_coercion_failure() {
        let workspace = tempfile::tempdir().unwrap();
        let executor = TaskExecutor::new(
            Arc::new(AgentRegistry::new()),
            Arc::new(SchemaFailClient),
            workspace.path().to_string_lossy().to_string(),
        );
        let spec = AgentStepSpec::new("step-x", "general", "assess", "Assess the thing.")
            .with_output_schema(verdict_schema());

        let outcome = executor.execute_step(spec, None).await;

        assert!(
            !outcome.success,
            "a step whose output can't satisfy the schema is demoted to failure"
        );
        assert_eq!(outcome.structured, None, "no validated object on failure");
        assert!(
            outcome.output.contains("[structured output failed"),
            "the demotion marker is appended: {}",
            outcome.output
        );
    }

    #[tokio::test]
    async fn parallel_isolates_schema_coercion_failure_from_sibling() {
        let workspace = tempfile::tempdir().unwrap();
        let executor: Arc<dyn AgentExecutor> = Arc::new(TaskExecutor::new(
            Arc::new(AgentRegistry::new()),
            Arc::new(SchemaFailClient),
            workspace.path().to_string_lossy().to_string(),
        ));
        // A plain step (no schema → succeeds) alongside a schema'd step whose
        // coercion fails. The failure must not drop or fail the sibling.
        let specs = vec![
            AgentStepSpec::new("plain", "general", "d", "p"),
            AgentStepSpec::new("schemad", "general", "d", "p").with_output_schema(verdict_schema()),
        ];
        let out = crate::orchestration::execute_steps_parallel(executor, specs, None).await;

        assert_eq!(out.len(), 2);
        assert_eq!(out[0].task_id, "plain");
        assert!(out[0].success, "no-schema sibling unaffected");
        assert_eq!(out[0].structured, None);
        assert_eq!(out[1].task_id, "schemad");
        assert!(!out[1].success, "schema-failing step surfaces as failure");
        assert_eq!(out[1].structured, None);
        assert!(out[1].output.contains("[structured output failed"));
    }

    #[tokio::test]
    async fn failed_step_with_schema_skips_coercion() {
        let workspace = tempfile::tempdir().unwrap();
        let executor = TaskExecutor::new(
            Arc::new(AgentRegistry::new()),
            Arc::new(SchemaCoercionClient),
            workspace.path().to_string_lossy().to_string(),
        );
        // Unknown agent → the run fails BEFORE coercion. The failure is the
        // run error, not a coercion failure — coercion must not run.
        let spec = AgentStepSpec::new("step-y", "no-such-agent", "d", "p")
            .with_output_schema(verdict_schema());

        let outcome = executor.execute_step(spec, None).await;

        assert!(!outcome.success);
        assert_eq!(outcome.structured, None);
        assert!(
            !outcome.output.contains("[structured output failed"),
            "coercion never ran — failure is the run error, not a coercion failure: {}",
            outcome.output
        );
    }

    struct StaticLlmClient {
        text: String,
    }

    impl StaticLlmClient {
        fn new(text: impl Into<String>) -> Self {
            Self { text: text.into() }
        }
    }

    #[async_trait::async_trait]
    impl LlmClient for StaticLlmClient {
        async fn complete(
            &self,
            messages: &[Message],
            system: Option<&str>,
            _tools: &[ToolDefinition],
        ) -> Result<LlmResponse> {
            if system == Some(crate::prompts::PRE_ANALYSIS_SYSTEM) {
                return Ok(pre_analysis_response(messages));
            }
            Ok(text_response(self.text.clone()))
        }

        async fn complete_streaming(
            &self,
            _messages: &[Message],
            _system: Option<&str>,
            _tools: &[ToolDefinition],
            _cancel_token: tokio_util::sync::CancellationToken,
        ) -> Result<mpsc::Receiver<StreamEvent>> {
            anyhow::bail!("streaming is not used by task executor tests")
        }
    }

    struct ConcurrentLlmClient {
        barrier: Arc<Barrier>,
        active: AtomicUsize,
        max_active: AtomicUsize,
    }

    impl ConcurrentLlmClient {
        fn new(task_count: usize) -> Self {
            Self {
                barrier: Arc::new(Barrier::new(task_count)),
                active: AtomicUsize::new(0),
                max_active: AtomicUsize::new(0),
            }
        }

        fn max_active(&self) -> usize {
            self.max_active.load(Ordering::SeqCst)
        }

        fn record_active(&self) {
            let active = self.active.fetch_add(1, Ordering::SeqCst) + 1;
            let mut observed = self.max_active.load(Ordering::SeqCst);
            while active > observed {
                match self.max_active.compare_exchange(
                    observed,
                    active,
                    Ordering::SeqCst,
                    Ordering::SeqCst,
                ) {
                    Ok(_) => break,
                    Err(next) => observed = next,
                }
            }
        }
    }

    struct LimitedConcurrencyLlmClient {
        active: AtomicUsize,
        max_active: AtomicUsize,
    }

    impl LimitedConcurrencyLlmClient {
        fn new() -> Self {
            Self {
                active: AtomicUsize::new(0),
                max_active: AtomicUsize::new(0),
            }
        }

        fn max_active(&self) -> usize {
            self.max_active.load(Ordering::SeqCst)
        }

        fn record_active(&self) {
            let active = self.active.fetch_add(1, Ordering::SeqCst) + 1;
            self.max_active.fetch_max(active, Ordering::SeqCst);
        }
    }

    #[async_trait::async_trait]
    impl LlmClient for LimitedConcurrencyLlmClient {
        async fn complete(
            &self,
            messages: &[Message],
            system: Option<&str>,
            _tools: &[ToolDefinition],
        ) -> Result<LlmResponse> {
            if system == Some(crate::prompts::PRE_ANALYSIS_SYSTEM) {
                return Ok(pre_analysis_response(messages));
            }

            let prompt = last_text(messages);
            self.record_active();
            tokio::time::sleep(Duration::from_millis(40)).await;
            self.active.fetch_sub(1, Ordering::SeqCst);
            Ok(text_response(format!("completed: {prompt}")))
        }

        async fn complete_streaming(
            &self,
            _messages: &[Message],
            _system: Option<&str>,
            _tools: &[ToolDefinition],
            _cancel_token: tokio_util::sync::CancellationToken,
        ) -> Result<mpsc::Receiver<StreamEvent>> {
            anyhow::bail!("streaming is not used by task executor tests")
        }
    }

    #[async_trait::async_trait]
    impl LlmClient for ConcurrentLlmClient {
        async fn complete(
            &self,
            messages: &[Message],
            system: Option<&str>,
            _tools: &[ToolDefinition],
        ) -> Result<LlmResponse> {
            if system == Some(crate::prompts::PRE_ANALYSIS_SYSTEM) {
                return Ok(pre_analysis_response(messages));
            }

            let prompt = last_text(messages);
            self.record_active();
            self.barrier.wait().await;
            if prompt.contains("slow") {
                tokio::time::sleep(Duration::from_millis(120)).await;
            } else {
                tokio::time::sleep(Duration::from_millis(10)).await;
            }
            self.active.fetch_sub(1, Ordering::SeqCst);
            Ok(text_response(format!("completed: {prompt}")))
        }

        async fn complete_streaming(
            &self,
            _messages: &[Message],
            _system: Option<&str>,
            _tools: &[ToolDefinition],
            _cancel_token: tokio_util::sync::CancellationToken,
        ) -> Result<mpsc::Receiver<StreamEvent>> {
            anyhow::bail!("streaming is not used by task executor tests")
        }
    }

    fn test_registry_with_writer() -> Arc<AgentRegistry> {
        let registry = AgentRegistry::new();
        let spec = crate::subagent::WorkerAgentSpec::custom("writer", "Write files")
            .with_permissions(PermissionPolicy::new().allow("write(*)").allow("read(*)"))
            .with_prompt("Write files when asked.")
            .with_max_steps(3);
        registry.register(spec.into_agent_definition());
        Arc::new(registry)
    }

    fn test_registry_with_text_worker() -> Arc<AgentRegistry> {
        let registry = AgentRegistry::new();
        let spec = crate::subagent::WorkerAgentSpec::custom("worker", "Text worker")
            .with_prompt("Return a concise result.")
            .with_max_steps(1);
        registry.register(spec.into_agent_definition());
        Arc::new(registry)
    }

    #[tokio::test]
    async fn task_child_run_permission_allow() {
        let workspace = tempfile::tempdir().unwrap();
        let mock = Arc::new(MockLlmClient::new(vec![
            MockLlmClient::tool_call_response(
                "t1",
                "write",
                serde_json::json!({
                    "file_path": workspace.path().join("out.txt").to_string_lossy(),
                    "content": "WRITTEN"
                }),
            ),
            MockLlmClient::text_response("Done."),
        ]));

        let executor = TaskExecutor::new(
            test_registry_with_writer(),
            mock,
            workspace.path().to_string_lossy().to_string(),
        );

        let result = executor
            .execute(
                TaskParams {
                    agent: "writer".to_string(),
                    description: "Write file".to_string(),
                    prompt: "Write out.txt".to_string(),
                    background: false,
                    max_steps: Some(3),
                },
                None,
                None,
            )
            .await
            .unwrap();

        assert!(
            result.success,
            "child run should succeed: {}",
            result.output
        );
        assert!(
            !result.output.contains("Permission denied"),
            "no permission denial: {}",
            result.output
        );
        let content = std::fs::read_to_string(workspace.path().join("out.txt")).unwrap();
        assert_eq!(content, "WRITTEN");
    }

    #[tokio::test]
    async fn task_child_run_permission_deny() {
        let workspace = tempfile::tempdir().unwrap();
        let registry = AgentRegistry::new();
        let spec = crate::subagent::WorkerAgentSpec::custom("restricted", "Restricted agent")
            .with_permissions(PermissionPolicy::new().allow("read(*)").deny("bash(*)"))
            .with_max_steps(3);
        registry.register(spec.into_agent_definition());

        let mock = Arc::new(MockLlmClient::new(vec![
            MockLlmClient::tool_call_response(
                "t1",
                "bash",
                serde_json::json!({"command": "echo hello"}),
            ),
            MockLlmClient::text_response("Could not run bash."),
        ]));

        let executor = TaskExecutor::new(
            Arc::new(registry),
            mock,
            workspace.path().to_string_lossy().to_string(),
        );

        let result = executor
            .execute(
                TaskParams {
                    agent: "restricted".to_string(),
                    description: "Try bash".to_string(),
                    prompt: "Run echo hello".to_string(),
                    background: false,
                    max_steps: Some(3),
                },
                None,
                None,
            )
            .await
            .unwrap();

        // The agent completes (LLM responds after denial), but bash was denied.
        // The denial is sent as a tool result to the LLM, which then responds.
        assert!(result.success, "agent should complete: {}", result.output);
    }

    #[tokio::test]
    async fn task_child_run_confirmation_auto_approve() {
        let workspace = tempfile::tempdir().unwrap();
        let registry = AgentRegistry::new();
        // Agent with allow("read(*)") — write is not in allow list, so it returns Ask.
        // With AutoApproveConfirmation (default for agents with permissions), Ask → approve.
        let spec = crate::subagent::WorkerAgentSpec::custom("reader-writer", "Read and write")
            .with_permissions(PermissionPolicy::new().allow("read(*)"))
            .with_max_steps(3);
        registry.register(spec.into_agent_definition());

        let mock = Arc::new(MockLlmClient::new(vec![
            MockLlmClient::tool_call_response(
                "t1",
                "write",
                serde_json::json!({
                    "file_path": workspace.path().join("auto.txt").to_string_lossy(),
                    "content": "AUTO_APPROVED"
                }),
            ),
            MockLlmClient::text_response("Written."),
        ]));

        let executor = TaskExecutor::new(
            Arc::new(registry),
            mock,
            workspace.path().to_string_lossy().to_string(),
        );

        let result = executor
            .execute(
                TaskParams {
                    agent: "reader-writer".to_string(),
                    description: "Write via auto-approve".to_string(),
                    prompt: "Write auto.txt".to_string(),
                    background: false,
                    max_steps: Some(3),
                },
                None,
                None,
            )
            .await
            .unwrap();

        assert!(
            result.success,
            "Ask should be auto-approved: {}",
            result.output
        );
        assert!(
            !result.output.contains("MissingConfirmationManager"),
            "no MissingConfirmationManager: {}",
            result.output
        );
    }

    #[tokio::test]
    async fn task_child_run_step_budget_enforced() {
        let workspace = tempfile::tempdir().unwrap();
        let mock = Arc::new(MockLlmClient::new(vec![
            MockLlmClient::tool_call_response(
                "t1",
                "read",
                serde_json::json!({"file_path": "/tmp/a.txt"}),
            ),
            MockLlmClient::tool_call_response(
                "t2",
                "read",
                serde_json::json!({"file_path": "/tmp/b.txt"}),
            ),
            MockLlmClient::tool_call_response(
                "t3",
                "read",
                serde_json::json!({"file_path": "/tmp/c.txt"}),
            ),
            MockLlmClient::text_response("Should not reach here."),
        ]));

        let executor = TaskExecutor::new(
            test_registry_with_writer(),
            mock,
            workspace.path().to_string_lossy().to_string(),
        );

        let result = executor
            .execute(
                TaskParams {
                    agent: "writer".to_string(),
                    description: "Exceed budget".to_string(),
                    prompt: "Read many files".to_string(),
                    background: false,
                    max_steps: Some(2),
                },
                None,
                None,
            )
            .await
            .unwrap();

        // The agent should fail after exceeding 2 tool rounds
        assert!(
            !result.success,
            "should fail when exceeding step budget: {}",
            result.output
        );
        assert!(
            result.output.contains("Max tool rounds") || result.output.contains("max tool rounds"),
            "error should mention tool rounds: {}",
            result.output
        );
    }

    #[tokio::test]
    async fn parallel_task_executor_runs_children_concurrently_and_preserves_input_order() {
        let workspace = tempfile::tempdir().unwrap();
        let client = Arc::new(ConcurrentLlmClient::new(2));
        let executor = Arc::new(TaskExecutor::new(
            test_registry_with_text_worker(),
            client.clone(),
            workspace.path().to_string_lossy().to_string(),
        ));

        let tasks = vec![
            TaskParams {
                agent: "worker".to_string(),
                description: "Slow task".to_string(),
                prompt: "slow branch".to_string(),
                background: false,
                max_steps: Some(1),
            },
            TaskParams {
                agent: "worker".to_string(),
                description: "Fast task".to_string(),
                prompt: "fast branch".to_string(),
                background: false,
                max_steps: Some(1),
            },
        ];

        let results = tokio::time::timeout(
            Duration::from_secs(2),
            executor.execute_parallel(tasks, None, None),
        )
        .await
        .expect("parallel children should reach the barrier and complete");

        assert_eq!(results.len(), 2);
        assert!(
            client.max_active() >= 2,
            "expected concurrent child execution, max_active={}",
            client.max_active()
        );
        assert!(results[0].success);
        assert!(results[0].output.contains("slow branch"));
        assert!(results[1].success);
        assert!(results[1].output.contains("fast branch"));
    }

    #[tokio::test]
    async fn parallel_task_executor_respects_configured_concurrency_limit() {
        let workspace = tempfile::tempdir().unwrap();
        let client = Arc::new(LimitedConcurrencyLlmClient::new());
        let executor = Arc::new(
            TaskExecutor::new(
                test_registry_with_text_worker(),
                client.clone(),
                workspace.path().to_string_lossy().to_string(),
            )
            .with_max_parallel_tasks(2),
        );

        let tasks = (0..5)
            .map(|idx| TaskParams {
                agent: "worker".to_string(),
                description: format!("Task {idx}"),
                prompt: format!("branch {idx}"),
                background: false,
                max_steps: Some(1),
            })
            .collect::<Vec<_>>();

        let results = executor.execute_parallel(tasks, None, None).await;

        assert_eq!(results.len(), 5);
        assert!(results.iter().all(|result| result.success));
        assert_eq!(client.max_active(), 2);
    }

    #[tokio::test]
    async fn parallel_task_executor_isolates_unknown_agent_failure() {
        let workspace = tempfile::tempdir().unwrap();
        let executor = Arc::new(TaskExecutor::new(
            test_registry_with_text_worker(),
            Arc::new(StaticLlmClient::new("valid branch done")),
            workspace.path().to_string_lossy().to_string(),
        ));

        let tasks = vec![
            TaskParams {
                agent: "missing-agent".to_string(),
                description: "Missing".to_string(),
                prompt: "should fail".to_string(),
                background: false,
                max_steps: Some(1),
            },
            TaskParams {
                agent: "worker".to_string(),
                description: "Valid".to_string(),
                prompt: "should succeed".to_string(),
                background: false,
                max_steps: Some(1),
            },
        ];

        let results = executor.execute_parallel(tasks, None, None).await;

        assert_eq!(results.len(), 2);
        assert!(!results[0].success);
        assert_eq!(results[0].agent, "missing-agent");
        assert!(results[0].output.contains("Unknown agent type"));
        assert!(results[1].success);
        assert_eq!(results[1].agent, "worker");
        assert!(results[1].output.contains("valid branch done"));
    }

    #[tokio::test]
    async fn parallel_task_executor_emits_subagent_events_for_each_child() {
        let workspace = tempfile::tempdir().unwrap();
        let executor = Arc::new(TaskExecutor::new(
            test_registry_with_text_worker(),
            Arc::new(StaticLlmClient::new("done")),
            workspace.path().to_string_lossy().to_string(),
        ));
        let (tx, mut rx) = broadcast::channel(64);

        let tasks = vec![
            TaskParams {
                agent: "worker".to_string(),
                description: "One".to_string(),
                prompt: "first".to_string(),
                background: false,
                max_steps: Some(1),
            },
            TaskParams {
                agent: "worker".to_string(),
                description: "Two".to_string(),
                prompt: "second".to_string(),
                background: false,
                max_steps: Some(1),
            },
        ];

        let results = executor.execute_parallel(tasks, Some(tx), None).await;
        assert_eq!(results.len(), 2);
        tokio::time::sleep(Duration::from_millis(20)).await;

        let mut starts = Vec::new();
        let mut ends = Vec::new();
        let mut progress_statuses: Vec<String> = Vec::new();
        while let Ok(event) = rx.try_recv() {
            match event {
                AgentEvent::SubagentStart { description, .. } => starts.push(description),
                AgentEvent::SubagentEnd { agent, success, .. } => ends.push((agent, success)),
                AgentEvent::SubagentProgress { status, .. } => progress_statuses.push(status),
                _ => {}
            }
        }

        starts.sort();
        assert_eq!(starts, vec!["One".to_string(), "Two".to_string()]);
        assert_eq!(ends.len(), 2);
        assert!(ends
            .iter()
            .all(|(agent, success)| agent == "worker" && *success));
        // Each child loop emits at least one TurnEnd, so we expect at least
        // two synthesized turn_completed progress events across the run.
        assert!(
            progress_statuses
                .iter()
                .filter(|s| s == &"turn_completed")
                .count()
                >= 2,
            "expected at least two turn_completed progress events, got {:?}",
            progress_statuses
        );
    }

    #[tokio::test]
    async fn parallel_task_tool_reports_error_when_any_child_fails() {
        let workspace = tempfile::tempdir().unwrap();
        let executor = Arc::new(TaskExecutor::new(
            test_registry_with_text_worker(),
            Arc::new(StaticLlmClient::new("valid branch done")),
            workspace.path().to_string_lossy().to_string(),
        ));
        let tool = ParallelTaskTool::new(executor);
        let ctx = ToolContext::new(workspace.path().to_path_buf());

        let output = tool
            .execute(
                &serde_json::json!({
                    "tasks": [
                        {
                            "agent": "missing-agent",
                            "description": "Missing",
                            "prompt": "should fail"
                        },
                        {
                            "agent": "worker",
                            "description": "Valid",
                            "prompt": "should succeed"
                        }
                    ]
                }),
                &ctx,
            )
            .await
            .unwrap();

        assert!(
            !output.success,
            "parallel_task should fail when any child result fails"
        );
        assert!(output.content.contains("[ERR]"));
        assert!(output.content.contains("[OK]"));
        let metadata = output.metadata.expect("metadata");
        assert_eq!(metadata["task_count"], 2);
        assert_eq!(metadata["results"][0]["success"], false);
        assert_eq!(metadata["results"][1]["success"], true);
    }

    #[tokio::test]
    async fn parallel_task_both_inherit_permissions() {
        let workspace = tempfile::tempdir().unwrap();
        let mock = Arc::new(MockLlmClient::new(vec![
            // Task 1 responses
            MockLlmClient::tool_call_response(
                "t1",
                "write",
                serde_json::json!({
                    "file_path": workspace.path().join("p1.txt").to_string_lossy(),
                    "content": "P1"
                }),
            ),
            MockLlmClient::text_response("Done 1."),
            // Task 2 responses
            MockLlmClient::tool_call_response(
                "t2",
                "write",
                serde_json::json!({
                    "file_path": workspace.path().join("p2.txt").to_string_lossy(),
                    "content": "P2"
                }),
            ),
            MockLlmClient::text_response("Done 2."),
        ]));

        let executor = Arc::new(TaskExecutor::new(
            test_registry_with_writer(),
            mock,
            workspace.path().to_string_lossy().to_string(),
        ));

        let tasks = vec![
            TaskParams {
                agent: "writer".to_string(),
                description: "Write p1".to_string(),
                prompt: "Write p1.txt".to_string(),
                background: false,
                max_steps: Some(3),
            },
            TaskParams {
                agent: "writer".to_string(),
                description: "Write p2".to_string(),
                prompt: "Write p2.txt".to_string(),
                background: false,
                max_steps: Some(3),
            },
        ];

        let results = executor.execute_parallel(tasks, None, None).await;
        assert_eq!(results.len(), 2);

        for result in &results {
            assert!(
                result.success,
                "parallel child should succeed: {}",
                result.output
            );
        }
    }

    #[test]
    fn synthesize_progress_emits_tool_completed_for_tool_end() {
        let event = AgentEvent::ToolEnd {
            id: "call-1".to_string(),
            name: "bash".to_string(),
            output: "hello".to_string(),
            exit_code: 0,
            metadata: None,
            error_kind: None,
        };
        let progress =
            synthesize_subagent_progress(&event, "task-1", "task-run-task-1").expect("some");
        match progress {
            AgentEvent::SubagentProgress {
                task_id,
                session_id,
                status,
                metadata,
            } => {
                assert_eq!(task_id, "task-1");
                assert_eq!(session_id, "task-run-task-1");
                assert_eq!(status, "tool_completed");
                assert_eq!(metadata["tool"], "bash");
                assert_eq!(metadata["exit_code"], 0);
                assert_eq!(metadata["output_bytes"], 5);
                assert!(metadata.get("error_kind").is_none());
            }
            other => panic!("expected SubagentProgress, got {:?}", other),
        }
    }

    #[test]
    fn synthesize_progress_includes_error_kind_when_present() {
        let event = AgentEvent::ToolEnd {
            id: "call-2".to_string(),
            name: "edit".to_string(),
            output: "boom".to_string(),
            exit_code: 1,
            metadata: None,
            error_kind: Some(crate::tools::ToolErrorKind::NotFound {
                path: "missing.txt".to_string(),
            }),
        };
        let progress =
            synthesize_subagent_progress(&event, "task-x", "task-run-task-x").expect("some");
        if let AgentEvent::SubagentProgress { metadata, .. } = progress {
            assert!(
                metadata.get("error_kind").is_some(),
                "error_kind should propagate into metadata"
            );
        } else {
            panic!("expected SubagentProgress");
        }
    }

    #[test]
    fn synthesize_progress_emits_turn_completed_for_turn_end() {
        let event = AgentEvent::TurnEnd {
            turn: 3,
            usage: crate::llm::TokenUsage {
                prompt_tokens: 100,
                completion_tokens: 25,
                total_tokens: 125,
                cache_read_tokens: None,
                cache_write_tokens: None,
            },
        };
        let progress =
            synthesize_subagent_progress(&event, "task-1", "task-run-task-1").expect("some");
        if let AgentEvent::SubagentProgress {
            status, metadata, ..
        } = progress
        {
            assert_eq!(status, "turn_completed");
            assert_eq!(metadata["turn"], 3);
            assert_eq!(metadata["total_tokens"], 125);
            assert_eq!(metadata["prompt_tokens"], 100);
            assert_eq!(metadata["completion_tokens"], 25);
        } else {
            panic!("expected SubagentProgress");
        }
    }

    #[test]
    fn synthesize_progress_ignores_unrelated_events() {
        let ignored = [
            AgentEvent::TextDelta {
                text: "hi".to_string(),
            },
            AgentEvent::ToolStart {
                id: "x".to_string(),
                name: "bash".to_string(),
            },
            AgentEvent::TurnStart { turn: 1 },
            AgentEvent::SubagentStart {
                task_id: "nested".to_string(),
                session_id: "nested-run".to_string(),
                parent_session_id: "parent".to_string(),
                agent: "explore".to_string(),
                description: "nested".to_string(),
                started_ms: 0,
            },
        ];
        for event in &ignored {
            assert!(
                synthesize_subagent_progress(event, "task", "session").is_none(),
                "{:?} should not emit progress",
                event
            );
        }
    }
}