agent_engine/tools/subagent/

start.rs

//! SubagentStartTool — dispatch a reactive subagent and return a handle_id immediately.
//!
//! Unlike the one-shot `subagent` tool, this tool returns *before* the subagent
//! finishes. The caller gets a `handle_id` they can poll via `subagent_status`,
//! steer via `subagent_steer`, or block on via `subagent_collect`.

use serde_json::{json, Value};
use std::sync::atomic::Ordering;
use std::sync::{Arc, RwLock};
use tokio::sync::{mpsc, oneshot};

use super::super::{resolve_agent_prompt, Tool, ToolContext, NEXT_SUBAGENT_ID};
use crate::runtime::subagent::{SubagentHandle, SubagentResult, SubagentState, SubagentStatus};
use crate::{AgentEvent, Result, RuntimeError};

pub struct SubagentStartTool;

#[async_trait::async_trait]
impl Tool for SubagentStartTool {
    fn origin(&self) -> crate::tools::ToolOrigin {
        crate::tools::ToolOrigin::Builtin
    }

    fn name(&self) -> &str {
        "subagent_start"
    }

    fn description(&self) -> &str {
        "Dispatch a reactive subagent and return immediately with a handle_id. \
         The subagent runs in the background and is REACTIVE: when it finishes, a \
         completion event is pushed into your queue and you are woken with its \
         status — do NOT sleep, busy-wait, or poll in a loop waiting for it. After \
         dispatching, end your turn or keep doing other work; on the completion \
         event call subagent_collect once with reconciled=true. subagent_status is \
         for an occasional progress peek, and subagent_steer injects guidance \
         mid-run. Use this for parallel execution or when you want to continue \
         working while the subagent runs. For simple sequential delegation that \
         must block, use subagent instead. Provide either an agent name (resolves \
         from ~/.synaps-cli/agents/<name>.md) or a system_prompt string directly; \
         give an inline subagent a short name so it is recognisable in the \
         subagent panel."
    }

    fn parameters(&self) -> Value {
        json!({
            "type": "object",
            "properties": {
                "agent": {
                    "type": "string",
                    "description": "Agent name — resolves to ~/.synaps-cli/agents/<name>.md. Mutually exclusive with system_prompt."
                },
                "system_prompt": {
                    "type": "string",
                    "description": "Inline system prompt for the subagent. Use when you don't have a named agent file."
                },
                "name": {
                    "type": "string",
                    "description": "Short display name for an inline (system_prompt) subagent, e.g. \"gif-recorder\". Shown in the subagent panel and in status/collect output. Letters, digits, - and _ only, max 32. Ignored when agent is set; defaults to \"inline\"."
                },
                "task": {
                    "type": "string",
                    "description": "The task/prompt to send to the subagent."
                },
                "model": {
                    "type": "string",
                    "description": "Omit to inherit the session foreground qualified identity. Explicit values must be one of subagent_models' listed exact choices."
                },
                "role": {
                    "type": "string",
                    "enum": ["planner", "implementer", "tester", "reviewer", "researcher", "debugger"],
                    "description": "Typed orchestration role."
                },
                "write_policy": {
                    "oneOf": [
                        {"type": "object", "properties": {"mode": {"const": "read_only"}}, "required": ["mode"]},
                        {"type": "object", "properties": {"mode": {"const": "isolated_worktree"}}, "required": ["mode"]},
                        {"type": "object", "properties": {"mode": {"const": "non_overlapping_paths"}, "scopes": {"type": "array", "items": {"type": "string"}}}, "required": ["mode", "scopes"]}
                    ],
                    "description": "Declared write isolation and path scopes."
                },
                "timeout": {
                    "type": "integer",
                    "description": "Wall-clock limit in seconds. Omit for the session default (config `subagent_timeout`); 0 = no limit. A timed-out worker keeps its full conversation: subagent_resume continues it where it stopped."
                }
            },
            "required": ["task"]
        })
    }

    async fn execute(&self, params: Value, ctx: ToolContext) -> Result<String> {
        // ── Parse params ───────────────────────────────────────────────────────
        let task = params["task"]
            .as_str()
            .ok_or_else(|| RuntimeError::Tool("Missing 'task' parameter".to_string()))?
            .to_string();

        // Treat blank / whitespace / control-char strings as absent.
        // Some model providers serialize "unset" as "" rather than omitting the field,
        // and we must not try to resolve "" (or " ", or "\u{0}") as an agent name —
        // doing so produces a hard "Agent '' not found" error that the model then
        // retries forever with sentinel values instead of falling back to system_prompt.
        let is_blank = |s: &String| s.chars().all(|c| c.is_whitespace() || c.is_control());
        let agent_name = params["agent"]
            .as_str()
            .map(|s| s.to_string())
            .filter(|s| !is_blank(s));
        let inline_prompt = params["system_prompt"]
            .as_str()
            .map(|s| s.to_string())
            .filter(|s| !is_blank(s));
        let requested_model = params["model"].as_str();
        // Validate the registry before authorization, id allocation, or event emission;
        // the same borrow is reused below when publishing the authorized handle.
        let registry = ctx.capabilities.subagent_registry.as_ref().ok_or_else(|| {
            RuntimeError::Tool("subagent_start requires a subagent_registry in ToolContext".into())
        })?;
        let system_prompt = match (&agent_name, &inline_prompt) {
            (Some(name), _) => resolve_agent_prompt(name).map_err(RuntimeError::Tool)?,
            (None, Some(p)) => p.clone(),
            (None, None) => {
                return Err(RuntimeError::Tool(
                    "Must provide either 'agent' (name) or 'system_prompt' (inline). Got neither."
                        .to_string(),
                ));
            }
        };
        let subagent_id = NEXT_SUBAGENT_ID.fetch_add(1, Ordering::Relaxed);
        let handle_id = format!("sa_{}", subagent_id);
        let orchestration = ctx
            .capabilities
            .orchestration
            .as_ref()
            .ok_or_else(|| RuntimeError::Tool("delegation policy unavailable".into()))?;
        orchestration
            .reserve_delegation(&handle_id, ctx.capabilities.delegation_parent.as_deref())
            .map_err(|reason| RuntimeError::Tool(format!("delegation denied: {reason}")))?;
        let decision = orchestration
            .resolve_and_authorize(&handle_id, requested_model)
            .map_err(|error| {
                orchestration
                    .release_delegation(&handle_id, ctx.capabilities.delegation_parent.as_deref());
                RuntimeError::Tool(error.to_string())
            })?;
        let model = decision.model.as_str().to_owned();
        let codex_parent_plan = ctx.capabilities.codex_parent_plan.clone();
        let memory_backend = ctx.capabilities.memory_backend.clone();
        let session_allow_all = ctx.capabilities.session_allow_all.clone();
        let telemetry = ctx.capabilities.telemetry.clone();
        let timeout_secs = params["timeout"]
            .as_u64()
            .unwrap_or(ctx.limits.subagent_timeout);

        let label = super::subagent_label(agent_name.as_deref(), params["name"].as_str());
        let task_preview: String = task.chars().take(80).collect();
        let task_full = task.clone();

        tracing::info!(
            "subagent_start: dispatching '{}' (id={}) model={}",
            label,
            handle_id,
            model
        );

        // ── Shared state ───────────────────────────────────────────────────────
        let state = Arc::new(RwLock::new(SubagentState::new()));
        state.write().unwrap().archive_meta = Some(crate::runtime::subagent::SubagentArchiveMeta {
            agent_name: label.clone(),
            model: model.clone(),
            system_prompt: system_prompt.clone(),
            timeout_secs,
        });

        state.write().unwrap().archive_dir = Some(super::archive::process_dir());
        let initial_messages = vec![Arc::new(
            serde_json::json!({"role": "user", "content": task.clone()}),
        )];
        state.write().unwrap().history = Some(initial_messages.clone());

        // ── Channels ───────────────────────────────────────────────────────────
        let (steer_tx, steer_rx) = mpsc::unbounded_channel::<String>();
        let (shutdown_tx, shutdown_rx) = oneshot::channel::<()>();
        let (result_tx, result_rx) = oneshot::channel::<SubagentResult>();

        // ── Forward SubagentStart event to TUI ─────────────────────────────────
        if let Some(ref tx) = ctx.channels.tx_events {
            let _ = tx.send(crate::StreamEvent::Agent(AgentEvent::SubagentStart {
                subagent_id,
                agent_name: label.clone(),
                task_preview: task_preview.clone(),
            }));
        }

        // ── Clone state for the spawned thread ─────────────────────────────────
        let state_t = Arc::clone(&state);
        let task_full_a = task_full.clone();
        let label_inner = label.clone();
        let model_inner = model.clone();
        let tx_events_inner = ctx.channels.tx_events.clone();
        let start_time = std::time::Instant::now();
        let parent_queue = ctx.capabilities.event_queue.clone();
        let handle_id_inner = handle_id.clone();
        let child_parent_id = handle_id.clone();

        // ── Build and register handle BEFORE spawning ─────────────────────────
        // This closes the publish-before-register race: finalize_subagent can
        // push a completion event the instant the thread exits, but the registry
        // must already contain the handle so the parent's collect succeeds.
        let system_prompt_for_handle = system_prompt.clone();
        let handle = SubagentHandle::new(
            handle_id.clone(),
            subagent_id,
            label.clone(),
            task_preview,
            model.clone(),
            system_prompt_for_handle,
            timeout_secs,
            Arc::clone(&state),
            Some(steer_tx),
            Some(shutdown_tx),
            Some(result_rx),
        )
        .with_authorization(&decision);
        {
            let mut reg = registry.lock().unwrap();
            reg.register_with_cancellation(handle, ctx.capabilities.launch_cancel.as_ref());
        }

        let orchestration = ctx.capabilities.orchestration.as_ref().unwrap();
        let orchestration_for_worker = Arc::clone(orchestration);
        let parent_for_worker = ctx.capabilities.delegation_parent.clone();
        if let Err(error) = orchestration.mark_starting(&handle_id) {
            orchestration.rollback(&handle_id);
            orchestration
                .release_delegation(&handle_id, ctx.capabilities.delegation_parent.as_deref());
            return Err(RuntimeError::Tool(error));
        }

        // ── Spawn subagent thread (mirrors subagent.rs) ────────────────────────
        let thread_handle = std::thread::spawn(move || {
            // Pre-clone for finalizer — catch_unwind moves state_t and label_inner
            let state_for_finalizer = Arc::clone(&state_t);
            let label_for_finalizer = label_inner.clone();
            let orchestration_for_runtime = Arc::clone(&orchestration_for_worker);

            let mut pending_result = None;
            let panic_result = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
                let rt = match tokio::runtime::Builder::new_current_thread()
                    .enable_all()
                    .build()
                {
                    Ok(rt) => rt,
                    Err(_) => {
                        state_t.write().unwrap().status =
                            SubagentStatus::Failed("runtime initialization failed".into());
                        return;
                    }
                };

                // Clones for the async block — the outer closure still needs the originals.
                let state_a = Arc::clone(&state_t);
                let label_a = label_inner.clone();
                let model_a = model_inner.clone();
                let tx_events_a = tx_events_inner.clone();
                let task_for_timeout = task_full_a.clone();
                let task_for_complete = task_full_a;

                let outcome: std::result::Result<SubagentResult, String> =
                    rt.block_on(async move {
                        // Host-built worker (shared client/creds/token cache, cached
                        // registry) or the legacy fresh runtime — see `spawn_runtime`.
                        let mut runtime = match super::spawn_runtime().await {
                            Ok(r) => r,
                            Err(_) => return Err("subagent runtime initialization failed".into()),
                        };

                        // Apply subagent spawn policy: worker role, 5m cache TTL, worker
                        // turn budget. Subagents are short-lived one-shots — paying the 1h
                        // write premium (~2× input price) on them is unrecoverable waste
                        // (~$0.23 per 10-spawn fan-out). (#110)
                        super::apply_subagent_runtime_policy(
                            &mut runtime,
                            &crate::config::load_config(),
                            memory_backend.as_ref(),
                            session_allow_all.as_ref(),
                            telemetry.as_ref(),
                        );
                        runtime.set_system_prompt(super::compose_system_prompt(
                            system_prompt,
                            runtime.memory_backend_is_axel(),
                        ));
                        runtime.set_model(model_a.clone());
                        super::apply_codex_worker_reasoning(
                            &mut runtime,
                            codex_parent_plan.as_ref(),
                        );
                        runtime
                            .install_worker_orchestration(Arc::clone(&orchestration_for_runtime));
                        runtime.set_delegation_parent(Some(child_parent_id.clone()));

                        let cancel = crate::CancellationToken::new();
                        let cancel_inner = cancel.clone();
                        tokio::spawn(async move {
                            let _ = shutdown_rx.await;
                            cancel_inner.cancel();
                        });

                        let cancel_on_timeout = cancel.clone();
                        let stream = runtime.run_stream_with_messages(
                                initial_messages,
                                cancel,
                                Some(steer_rx),
                                None,
                                false,
                            ).await;

                        super::drive::WorkerDrive {
                            state: Arc::clone(&state_a),
                            tx_events: tx_events_a.clone(),
                            subagent_id,
                            label: label_a.clone(),
                            model: model_a.clone(),
                            timeout_secs,
                            cancel: cancel_on_timeout,
                            task: task_for_timeout,
                        }
                        .run(stream)
                        .await
                    });

                match outcome {
                    Ok(sa_result) => {
                        // Only overwrite Running → Completed (don't stomp TimedOut or a cancellation).
                        {
                            let mut s = state_t.write().unwrap();
                            if matches!(s.status, SubagentStatus::Running) && !s.cancel_requested {
                                s.status = SubagentStatus::Completed;
                                s.conversation_state = vec![
                                    serde_json::json!({"role": "user", "content": task_for_complete.clone()}),
                                    serde_json::json!({"role": "assistant", "content": sa_result.text.clone()}),
                                ];
                            }
                        }
                        pending_result = Some(sa_result);
                    }
                    Err(e) => {
                        state_t.write().unwrap().status = SubagentStatus::Failed(e);
                    }
                }
            }));

            if let Err(panic_info) = panic_result {
                let msg = if let Some(s) = panic_info.downcast_ref::<&str>() {
                    s.to_string()
                } else if let Some(s) = panic_info.downcast_ref::<String>() {
                    s.clone()
                } else {
                    "unknown panic".to_string()
                };
                tracing::error!("Subagent thread panicked: {}", msg);
                state_t.write().unwrap_or_else(|p| p.into_inner()).status =
                    SubagentStatus::Failed(format!("panic: {}", msg));
            }

            // ── Terminal finalizer — exactly once, outside catch_unwind ────────
            // Covers all paths: Ok, Err, timeout, panic, early tokio-build failure.
            super::finalize::finalize_subagent(
                &state_for_finalizer,
                parent_queue.as_ref(),
                &handle_id_inner,
                subagent_id,
                &label_for_finalizer,
                start_time,
                None, // start.rs: not a resume
            );
            // A cancelled worker publishes no wake, so nothing would ever
            // prompt a collect: retire it before any terminal publication.
            super::finalize::retire_if_cancelled(
                &state_for_finalizer,
                &orchestration_for_worker,
                &handle_id_inner,
            );
            // All terminal publications follow the recovery/finalization barrier.
            if let Some(ref tx) = tx_events_inner {
                let s = state_for_finalizer
                    .read()
                    .unwrap_or_else(|p| p.into_inner());
                let preview = match s.status.failure_reason() {
                    Some(reason) => format!("ERROR: {reason}"),
                    None => s.partial_text.chars().take(120).collect(),
                };
                let _ = tx.send(crate::StreamEvent::Agent(AgentEvent::SubagentDone {
                    subagent_id,
                    agent_name: label_for_finalizer.clone(),
                    result_preview: preview,
                    duration_secs: start_time.elapsed().as_secs_f64(),
                }));
            }
            if let Some(result) = pending_result {
                let _ = result_tx.send(result);
            }
            orchestration_for_worker
                .release_delegation(&handle_id_inner, parent_for_worker.as_deref());
        });

        // ── Attach the already-persisted handle to the started thread ──────────
        {
            let mut reg = registry.lock().unwrap();
            if let Some(h) = reg.get_mut(&handle_id) {
                h.set_thread_handle(thread_handle);
            }
        }
        if let Err(error) = ctx
            .capabilities
            .orchestration
            .as_ref()
            .unwrap()
            .mark_running(&handle_id)
        {
            let thread = {
                let mut reg = registry.lock().unwrap();
                if let Some(handle) = reg.get_mut(&handle_id) {
                    handle.cancel();
                }
                reg.remove(&handle_id)
            };
            if let Some(handle) = thread {
                let _ = handle.collect().await;
            }
            orchestration.rollback(&handle_id);
            orchestration
                .release_delegation(&handle_id, ctx.capabilities.delegation_parent.as_deref());
            return Err(RuntimeError::Tool(error));
        }
        // The thread may have been cancelled and finalized before it was
        // marked running; its own retire attempt was then too early.
        super::finalize::retire_if_cancelled(&state, orchestration, &handle_id);

        Ok(json!({
            "handle_id":  handle_id,
            "agent_name": label,
            "status":     "running"
        })
        .to_string())
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::tools::test_helpers::create_tool_context;
    use crate::tools::{SubagentRegistry, Tool};
    use serde_json::json;
    use std::sync::{Arc, Mutex};

    // V4: registry=None must return Err immediately, before any thread is spawned.
    #[tokio::test]
    async fn test_subagent_start_no_registry_returns_err() {
        let tool = SubagentStartTool;
        let ctx = create_tool_context(); // subagent_registry is None by default

        let params = json!({
            "system_prompt": "You are a test subagent.",
            "task": "Say ok",
        });

        let result = tool.execute(params, ctx).await;
        assert!(
            result.is_err(),
            "missing registry must return Err before spawn"
        );
        let msg = format!("{:?}", result.unwrap_err());
        assert!(
            msg.contains("subagent_registry"),
            "error must mention subagent_registry: {msg}"
        );
    }

    /// These tests launch a real worker thread. Run unpinned, it read the
    /// developer's real `auth.json` (and could reach the provider), and wrote
    /// `logs/subagents` and `subagent-history` under the real `~/.synaps-cli`,
    /// after the test returned. Pin an empty base dir and clear the
    /// environment's provider credentials/endpoints (`OfflineProviderEnv`), so
    /// the worker fails at auth, offline; keep both until it has exited.
    fn join_worker(registry: &Mutex<SubagentRegistry>, handle_id: &str) {
        let thread = registry
            .lock()
            .unwrap()
            .get_mut(handle_id)
            .and_then(|h| h.thread_handle.take());
        if let Some(thread) = thread {
            let _ = thread.join();
        }
    }

    #[tokio::test]
    #[serial_test::serial(synaps_base_dir)]
    async fn test_subagent_start_blank_agent_uses_system_prompt() {
        let _base = crate::test_env::BaseDirGuard::new();
        let _offline = crate::test_env::OfflineProviderEnv::new();
        let tool = SubagentStartTool;
        let mut ctx = create_tool_context();
        let registry = Arc::new(Mutex::new(SubagentRegistry::new()));
        ctx.capabilities.subagent_registry = Some(registry.clone());

        let params = json!({
            "agent": "",
            "system_prompt": "You are a concise test subagent. Reply with only: ok",
            "task": "Say ok",
            "model": "anthropic/claude-sonnet-4-6",
            "timeout": 1
        });

        let result = tool.execute(params, ctx).await;
        assert!(
            result.is_ok(),
            "blank agent should not be resolved as ~/.synaps-cli/agents/.md: {result:?}"
        );
        let body: serde_json::Value = serde_json::from_str(&result.unwrap()).unwrap();
        assert_eq!(body["agent_name"], "inline");
        let handle_id = body["handle_id"].as_str().unwrap_or_default();
        assert!(handle_id.starts_with("sa_"));
        join_worker(&registry, handle_id);
    }

    #[tokio::test]
    #[serial_test::serial(synaps_base_dir)]
    async fn test_subagent_start_inline_name_labels_the_handle() {
        let _base = crate::test_env::BaseDirGuard::new();
        let _offline = crate::test_env::OfflineProviderEnv::new();
        let tool = SubagentStartTool;
        let mut ctx = create_tool_context();
        let registry = Arc::new(Mutex::new(SubagentRegistry::new()));
        ctx.capabilities.subagent_registry = Some(registry.clone());

        let params = json!({
            "system_prompt": "You are a concise test subagent. Reply with only: ok",
            "name": "gif recorder",
            "task": "Say ok",
            "model": "anthropic/claude-sonnet-4-6",
            "timeout": 1
        });

        let result = tool.execute(params, ctx).await.expect("dispatch");
        let body: serde_json::Value = serde_json::from_str(&result).unwrap();
        assert_eq!(body["agent_name"], "gif-recorder");
        let handle_id = body["handle_id"].as_str().unwrap();
        {
            let reg = registry.lock().unwrap();
            assert_eq!(
                reg.get(handle_id).expect("registered").agent_name,
                "gif-recorder"
            );
        }
        join_worker(&registry, handle_id);
    }
}