harness-tools + app: task tool, subagent spawner, board wiring (M4)

The `task` tool delegates to a SubagentSpawner (owned by the composition root):
resolves the agent, enforces the depth limit, applies permission intersection,
filters tools per-agent, and runs the child session foreground or background.
Foreground returns the child's final text; background registers on the job board
and detaches under the parent run token. The run loop injects the board into
primary-agent requests and reconciles terminal jobs each step.

Integration tests cover foreground run + alias reuse continuing the same child
session, depth-limit and unknown-agent rejection, and board prompt injection.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
2026-07-09 06:29:12 +02:00
co-authored by Claude Opus 4.8
parent 8c859d91c9
commit ecb3267a50
18 changed files with 1137 additions and 67 deletions
+156
View File
@@ -22,6 +22,8 @@ pub struct RunConfig {
pub instructions: Vec<String>,
/// Pricing for `model`, from models.dev metadata. `None` leaves cost at 0.
pub cost: Option<crate::types::ModelCost>,
/// Whether to append the background job board to requests (primary/delegating agents).
pub inject_job_board: bool,
}
/// Adds a step's usage/cost onto the persisted session and republishes it. Cost accounting is
@@ -198,6 +200,22 @@ pub async fn run_session(
wire_messages.extend(convert_message(message, &parts));
}
// Append the (synthetic, non-persisted) job board to the last user message so the
// orchestrator sees running/reusable subtasks. docs/04-multiagent.md.
if run_config.inject_job_board {
if let Some(board) = &ctx.job_board {
if let Some(block) = board.format_for_prompt() {
if let Some(last_user) = wire_messages
.iter_mut()
.rev()
.find(|m| m.role == WireRole::User)
{
last_user.content.push(WireContent::Text { text: block });
}
}
}
}
let system_blocks = system::assemble(
system::env_header(&ctx.cwd),
&run_config.agent_prompt,
@@ -260,6 +278,15 @@ pub async fn run_session(
}
Ok(outcome) => {
accumulate_session_usage(&ctx, &outcome.usage, outcome.cost, now_fn()).await;
// A completed step means the orchestrator has now seen any terminal jobs
// that were on the board this turn; mark them reconciled.
if run_config.inject_job_board {
if let Some(board) = &ctx.job_board {
if let Err(e) = board.reconcile_terminal(now_fn()).await {
tracing::warn!(error = %e, "failed to reconcile job board");
}
}
}
match outcome.result {
StepResult::Continue => continue,
StepResult::Stop => return RunOutcome::Stopped,
@@ -376,11 +403,14 @@ mod tests {
permissions,
static_rules: Vec::new(),
extra_rules: Arc::new(std::sync::Mutex::new(Vec::new())),
parent_rules: Vec::new(),
session_id,
cwd: cwd.clone(),
data_dir: cwd.join("tool-output"),
cancel: CancellationToken::new(),
now: 1,
spawner: None,
job_board: None,
}
}
@@ -462,6 +492,7 @@ mod tests {
output: 15.0,
..Default::default()
}),
inject_job_board: false,
};
let outcome = run_session(Arc::new(provider), ctx, &run_config, || 2).await;
@@ -569,6 +600,7 @@ mod tests {
max_steps: 10,
instructions: Vec::new(),
cost: None,
inject_job_board: false,
};
let outcome = run_session(Arc::new(provider), ctx, &run_config, || 2).await;
@@ -578,4 +610,128 @@ mod tests {
let messages = store.messages(session_id).await.unwrap();
assert_eq!(messages.len(), 1);
}
/// Records the last request it was asked to stream so tests can assert on prompt content.
struct CapturingProvider {
last: StdMutex<Option<LlmRequest>>,
}
#[async_trait]
impl Provider for CapturingProvider {
fn id(&self) -> &str {
"mock"
}
async fn list_models(&self) -> Result<Vec<ModelInfo>, ProviderError> {
Ok(vec![])
}
async fn stream(
&self,
req: LlmRequest,
_cancel: CancellationToken,
) -> Result<LlmEventStream, ProviderError> {
*self.last.lock().unwrap() = Some(req);
let events = vec![
Ok(LlmEvent::TextStart { id: "t".into() }),
Ok(LlmEvent::TextDelta {
id: "t".into(),
text: "ok".into(),
}),
Ok(LlmEvent::TextEnd { id: "t".into() }),
Ok(LlmEvent::Finish {
reason: FinishReason::Stop,
usage: usage(1, 1),
}),
];
Ok(Box::pin(futures::stream::iter(events)))
}
}
#[tokio::test]
async fn job_board_is_injected_into_the_last_user_message() {
use crate::engine::jobs::{JobBoard, LaunchSpec};
let store = Store::open_in_memory().unwrap();
let bus = EventBus::new();
let model = ModelRef::new("mock", "mock-model");
let session = Session::new_root("orchestrator", model.clone(), 1);
let session_id = session.id.clone();
store.upsert_session(session).await.unwrap();
let user_message = Message::new_user(session_id.clone(), 1);
store.upsert_message(user_message.clone()).await.unwrap();
store
.upsert_part(Part {
id: crate::types::PartId::new(),
message_id: user_message.id.clone(),
session_id: session_id.clone(),
idx: 0,
body: PartBody::Text {
text: "carry on".into(),
synthetic: false,
},
})
.await
.unwrap();
// A board with one running job for this session.
let board = std::sync::Arc::new(
JobBoard::load(store.clone(), bus.clone(), &session_id, 2)
.await
.unwrap(),
);
board
.register_launch(
LaunchSpec {
task_id: "t1".into(),
parent_session: session_id.clone(),
child_session: SessionId::new(),
agent: "explorer".into(),
description: "map auth".into(),
objective: Some("map the auth flow".into()),
},
1,
)
.await
.unwrap();
let cwd = tempfile::tempdir().unwrap();
let mut ctx = make_ctx(store, bus, session_id, cwd.path().to_path_buf()).await;
ctx.job_board = Some(board);
let run_config = RunConfig {
agent_name: "orchestrator".into(),
agent_prompt: "You orchestrate.".into(),
model,
temperature: None,
max_steps: 1,
instructions: Vec::new(),
cost: None,
inject_job_board: true,
};
let provider = std::sync::Arc::new(CapturingProvider {
last: StdMutex::new(None),
});
let outcome = run_session(provider.clone(), ctx, &run_config, || 2).await;
assert!(matches!(outcome, RunOutcome::Stopped));
let req = provider.last.lock().unwrap().clone().expect("a request");
let last_user = req
.messages
.iter()
.rev()
.find(|m| m.role == WireRole::User)
.expect("a user message");
let text: String = last_user
.content
.iter()
.filter_map(|c| match c {
WireContent::Text { text } => Some(text.as_str()),
_ => None,
})
.collect::<Vec<_>>()
.join("\n");
assert!(text.contains("Background Job Board"), "got: {text}");
assert!(text.contains("exp-1"), "got: {text}");
assert!(text.contains("map the auth flow"), "got: {text}");
}
}