Watch
0
0
Fork
You've already forked hyperhive
0

subagents publish their model and effort as headers on every row, and the swarm shows them

The subagent daemon stamps Swarm-Model and Swarm-Effort on each row it
publishes. The list endpoint reads them off the newest row it already fetches,
so a long-running subagent keeps them past the stream's 24 hour window, and the
terminal page's menu and the agent panel's list show them beside the name.
This commit is contained in:
iris 2026-10-28 20:54:16 +01:00 • committed by mara
commit da5fcbfd9b
10 changed files with 219 additions and 25 deletions

View file

@ -63,6 +63,9 @@ does for the agent. The list refreshes every 10 seconds, so a subagent the
agent spawns shows up once it writes output. A subagent that drops out of the
list closes its terminal, and a later subagent with the same name starts
unpicked.
Each subagent in the list shows the model and effort it runs, like
`opus · high`, beside its name; one on the default model shows only the effort.
The `/agents/<name>/terminal` page's subagent list shows them too.
The preview carries only the connection badge and **expand**, so it stays clear
of the text. What the agent is doing shows in the details table above it, in
@ -433,7 +436,12 @@ into a stream named `term-sub-<agent>` that swarm-controller creates for every
declared agent. The stream keeps rows for 24 hours. swarm-controller serves the list as
`GET /api/agents/<name>/subagents`, the subagents named by the subjects in that
stream, each with `last_row_unix`, the stream's publish time of its newest row,
or `null` when that row isn't read within two seconds. The lookups run
or `null` when that row isn't read within two seconds. Each row also carries
the `Swarm-Model` and `Swarm-Effort` message headers, the model and effort the
subagent runs with, and the list returns them from the newest row as
`model` and `effort`. A subagent on the default model has no `Swarm-Model`
header and lists `model: null`; every row has the headers, so a subagent that
runs past the 24-hour window still shows them. The lookups run
concurrently, sixteen at a time, so the list waits at most two seconds for
each sixteen subagents.
swarm-controller also relays one subagent's rows as SSE on

View file

@ -12,7 +12,7 @@ import { useEffect, useState } from "preact/hooks";
import { ApiErrorPanel } from "@hive/shared/api-error-panel.js";
import { Badge } from "@hive/shared/badge.js";
import { AgentTermPreview } from "./AgentTermPreview.js";
import { STALE_S, stillListed } from "./subagent-list.js";
import { STALE_S, runLabel, stillListed } from "./subagent-list.js";
import { useSubagentList } from "./useSubagentList.js";
import { RelativeTime } from "../../ui/relative-time/RelativeTime.js";
import "./SubagentTerms.css";
@ -41,7 +41,9 @@ export function SubagentTerms({ agentName }: { agentName: string }) {
<section class="ui-subagent-terms" aria-label={`${agentName}'s subagents`}>
<div class="ui-subagent-terms-list">
<span class="ui-subagent-terms-heading">subagents</span>
{subagents?.map(({ name, last_row_unix }) => {
{subagents?.map((sub) => {
const { name, last_row_unix } = sub;
const runs = runLabel(sub);
const stale =
last_row_unix !== null && nowS - last_row_unix > STALE_S;
return (
@ -60,6 +62,7 @@ export function SubagentTerms({ agentName }: { agentName: string }) {
}
onClick={() => setPicked(name === selected ? null : name)}
/>
{runs && <span class="ui-subagent-terms-age">{runs}</span>}
{last_row_unix !== null && (
<span class="ui-subagent-terms-age">
<RelativeTime epochMs={last_row_unix * 1000} />

View file

@ -45,3 +45,9 @@
.ui-subagents-menu-back {
color: var(--blue);
}
/* What it runs sits after the name, not centred between name and age. */
.ui-subagents-menu-run {
margin-right: auto;
margin-left: 0.6em;
}

View file

@ -17,7 +17,12 @@ import { Panel } from "../../ui/panel/Panel.js";
import { Popover } from "../../ui/popover/Popover.js";
import { RelativeTime } from "../../ui/relative-time/RelativeTime.js";
import { termPaths } from "./AgentTermPreview.js";
import { runningLabel, splitRunning, type Subagent } from "./subagent-list.js";
import {
runLabel,
runningLabel,
splitRunning,
type Subagent,
} from "./subagent-list.js";
import { useSubagentList } from "./useSubagentList.js";
import "./SubagentsMenu.css";
@ -41,6 +46,11 @@ function Entry({
aria-current={current ? "page" : undefined}
>
<span class="ui-subagents-menu-name">{sub.name}</span>
{runLabel(sub) && (
<span class="ui-subagents-menu-age ui-subagents-menu-run">
{runLabel(sub)}
</span>
)}
{sub.last_row_unix !== null && (
<span class="ui-subagents-menu-age">
<RelativeTime epochMs={sub.last_row_unix * 1000} />

View file

@ -8,6 +8,19 @@ export interface Subagent {
name: string;
/** Unix seconds of its newest terminal row, `null` when not readable. */
last_row_unix: number | null;
/** The model it runs; `null` on the harness's default or when not readable. */
model: string | null;
/** Its reasoning effort; `null` when not readable. */
effort: string | null;
}
/**
* What a subagent runs, as one short label: `opus · high`. Whichever of the
* two the controller could not read is left out, and an empty string means
* there is nothing to show.
*/
export function runLabel(sub: Pick<Subagent, "model" | "effort">): string {
return [sub.model, sub.effort].filter((v) => v).join(" · ");
}
/** A subagent quiet for longer than this is no longer counted as running. */

View file

@ -7,6 +7,7 @@ import {
STALE_S,
byRecency,
isRunning,
runLabel,
runningLabel,
splitRunning,
stillListed,
@ -16,6 +17,8 @@ const NOW = 1_000_000;
const sub = (name, ago) => ({
name,
last_row_unix: ago === null ? null : NOW - ago,
model: null,
effort: null,
});
test("a subagent with a recent row is running, a quiet one is not", () => {
@ -69,3 +72,11 @@ test("a pick holds only while the refreshed list still has it", () => {
assert.equal(stillListed("a", null), null);
assert.equal(stillListed(null, subs), null);
});
test("the run label names the model and effort, whichever are known", () => {
assert.equal(runLabel({ model: "opus", effort: "high" }), "opus · high");
assert.equal(runLabel({ model: "opus", effort: null }), "opus");
assert.equal(runLabel({ model: null, effort: "medium" }), "medium");
assert.equal(runLabel({ model: null, effort: null }), "");
assert.equal(runLabel({ model: "", effort: "" }), "");
});

View file

@ -1814,6 +1814,9 @@ fn continue_on_runtime(
struct LivenessSink {
state: Arc<State>,
name: String,
/// The model and effort this run was started with, stamped on every row
/// it publishes.
info: Arc<crate::swarm_term::RunInfo>,
/// Where to report the first non-terminal event, on a `resume` whose
/// caller is still waiting to hear whether it landed. `None` on a
/// `start`, which has nobody waiting.
@ -1845,11 +1848,20 @@ impl LivenessSink {
/// one to take it.
fn publish(&self, line: impl FnOnce() -> Output) {
if let Some(term) = &self.state.term {
term.offer(&self.name, line());
term.offer(&self.name, &self.info, line());
}
}
}
/// What a run's rows say about how it was started: its `config`'s model and
/// effort, which the swarm shows beside the subagent's name.
fn run_info(config: &Config) -> Arc<crate::swarm_term::RunInfo> {
Arc::new(crate::swarm_term::RunInfo {
model: config.model.clone(),
effort: config.effort.clone(),
})
}
/// Spawn the child (synchronous — `Claude::spawn` returns with a real pid
/// the instant the process exists), track it in `running`, and hand the
/// actual turn off to a background task so nobody blocks on the whole turn.
@ -1911,6 +1923,7 @@ fn spawn_and_track(
None
};
let info = run_info(config);
let config = config.clone();
let state = Arc::clone(state);
let task_name = name.to_owned();
@ -1922,6 +1935,7 @@ fn spawn_and_track(
let sink = LivenessSink {
state: Arc::clone(&state),
name: task_name.clone(),
info: Arc::clone(&info),
verdict: verdict.clone(),
};
let end = classify_end(running.wait(&prompt, &sink).await, searched.as_deref());
@ -2025,6 +2039,7 @@ fn spawn_and_track_acp(
state.clear_kill(name);
state.note_event(name);
let info = run_info(&config);
let state = Arc::clone(state);
let task_name = name.to_owned();
tokio::spawn(async move {
@ -2033,6 +2048,7 @@ fn spawn_and_track_acp(
let sink = LivenessSink {
state: Arc::clone(&state),
name: task_name.clone(),
info: Arc::clone(&info),
verdict: verdict.clone(),
};
let end = match runtime.run(&config, &prompt, &sink).await {
@ -3455,6 +3471,7 @@ mod tests {
let sink = LivenessSink {
state: Arc::clone(&state),
name: "n".to_owned(),
info: Arc::default(),
verdict: None,
};
(state, sink)
@ -3745,6 +3762,7 @@ mod tests {
let sink = LivenessSink {
state: Arc::new(State::new(PathBuf::from("/dev/null"), signal_url())),
name: "n".to_owned(),
info: Arc::default(),
verdict: Some(Arc::new(Mutex::new(Some(tx)))),
};
(rx, sink)

View file

@ -74,8 +74,32 @@ pub enum Output {
Stderr(String),
}
/// The model and effort a subagent was started with, published as headers on
/// each of its rows (`swarm_wire::subagent_term`). `None` is the harness's
/// default, which has no name to publish.
#[derive(Debug, Clone, Default, PartialEq, Eq)]
pub struct RunInfo {
pub model: Option<String>,
pub effort: Option<String>,
}
impl RunInfo {
/// The headers for one row: only the values there are.
fn headers(&self) -> async_nats::HeaderMap {
let mut headers = async_nats::HeaderMap::new();
if let Some(model) = self.model.as_deref().filter(|v| !v.is_empty()) {
headers.insert(subagent_term::HEADER_MODEL, model);
}
if let Some(effort) = self.effort.as_deref().filter(|v| !v.is_empty()) {
headers.insert(subagent_term::HEADER_EFFORT, effort);
}
headers
}
}
struct Line {
subagent: String,
info: Arc<RunInfo>,
output: Output,
/// Unix seconds when the sink saw the line.
ts: i64,
@ -131,14 +155,16 @@ impl Publisher {
(publisher, rx)
}
/// Queue `output` from `subagent` for publishing. Never blocks: a full
/// queue drops the line and counts it.
pub fn offer(&self, subagent: &str, output: Output) {
/// Queue `output` from `subagent` for publishing, with the `info` its rows
/// carry as headers. Never blocks: a full queue drops the line and counts
/// it.
pub fn offer(&self, subagent: &str, info: &Arc<RunInfo>, output: Output) {
let ts = SystemTime::now()
.duration_since(UNIX_EPOCH)
.map_or(0, |d| i64::try_from(d.as_secs()).unwrap_or(i64::MAX));
let line = Line {
subagent: subagent.to_owned(),
info: Arc::clone(info),
output,
ts,
};
@ -296,8 +322,9 @@ async fn run(mut rx: mpsc::Receiver<Line>, target: Target, dropped: Arc<AtomicU6
continue;
};
let subject = target.subject(&line.subagent);
let headers = line.info.headers();
for row in rows {
publish(&client, &subject, row).await;
publish(&client, &subject, &headers, row).await;
}
}
}
@ -314,7 +341,12 @@ fn classify(output: &Output, ts: i64, ctx: &mut ClassifyCtx) -> Vec<TermMsg> {
}
/// Offer one row. Every failure is terminal for that row and for nothing else.
async fn publish(client: &async_nats::Client, subject: &str, msg: TermMsg) {
async fn publish(
client: &async_nats::Client,
subject: &str,
headers: &async_nats::HeaderMap,
msg: TermMsg,
) {
// Unconnected, the client buffers rather than fails, and reports the
// library's default payload limit rather than the server's.
if let Err(e) = swarm_queue_client::ensure_connected(client) {
@ -338,7 +370,10 @@ async fn publish(client: &async_nats::Client, subject: &str, msg: TermMsg) {
return;
}
};
if let Err(e) = client.publish(subject.to_owned(), payload.into()).await {
let sent = client
.publish_with_headers(subject.to_owned(), headers.clone(), payload.into())
.await;
if let Err(e) = sent {
tracing::warn!(error = %e, "subagent terminal: publish failed, row dropped");
}
}
@ -412,11 +447,45 @@ mod tests {
#[test]
fn a_full_queue_drops_and_counts_rather_than_blocks() {
let (publisher, _rx) = Publisher::channel(1);
publisher.offer("scout", Output::Stdout("one".to_owned()));
publisher.offer("scout", Output::Stdout("two".to_owned()));
let info = Arc::new(RunInfo::default());
publisher.offer("scout", &info, Output::Stdout("one".to_owned()));
publisher.offer("scout", &info, Output::Stdout("two".to_owned()));
assert_eq!(publisher.dropped.load(Ordering::Relaxed), 1);
}
/// The list endpoint reads these two headers off a subagent's newest row.
#[test]
fn a_row_carries_the_model_and_effort_the_subagent_runs() {
let info = RunInfo {
model: Some("opus".to_owned()),
effort: Some("high".to_owned()),
};
let headers = info.headers();
assert_eq!(
headers
.get(subagent_term::HEADER_MODEL)
.map(ToString::to_string),
Some("opus".to_owned())
);
assert_eq!(
headers
.get(subagent_term::HEADER_EFFORT)
.map(ToString::to_string),
Some("high".to_owned())
);
}
/// A subagent on the default model names none, and an empty header would
/// read as a model called "".
#[test]
fn a_value_that_is_not_set_is_not_published() {
let info = RunInfo {
model: None,
effort: Some(String::new()),
};
assert!(info.headers().is_empty());
}
#[test]
fn every_kind_of_line_renders_as_stamped_rows() {
let mut ctx = ClassifyCtx::default();

View file

@ -11,6 +11,8 @@
//! holds rows for, read with `STREAM.INFO`. A missing stream lists nothing.
//! Each subagent carries the stream's own publish time of its newest row, so
//! a reader can tell a subagent that is running from one that has gone quiet.
//! The same row's `Swarm-Model` and `Swarm-Effort` headers
//! (`swarm_wire::subagent_term`) give the model and effort it runs with.
//!
//! **No agent-status lookup.** Unlike [`crate::term_stream`], an agent that
//! has never reported gets an empty stream, not a 404.
@ -120,6 +122,11 @@ pub(crate) struct SubagentEntry {
/// When its newest row in the stream was published, unix seconds, by the
/// stream's own clock. `None` when that row could not be read.
pub last_row_unix: Option<i64>,
/// The model it runs, as its newest row's `Swarm-Model` header says. `None`
/// when the row has none (the harness's default model) or could not be read.
pub model: Option<String>,
/// Its reasoning effort, from the `Swarm-Effort` header, with the same `None`.
pub effort: Option<String>,
}
#[utoipa::path(
@ -165,8 +172,8 @@ pub(crate) async fn list_subagents(
// item type trips a "`FnOnce` is not general enough" inference error where
// `routes!` checks this fn as a handler.
let (stream_ref, agent_ref) = (&stream, &agent);
let lookups: Vec<Result<i64, String>> = futures_util::stream::iter(names.clone())
.map(|name| async move { last_row_unix(stream_ref, agent_ref, &name).await })
let lookups: Vec<Result<NewestRow, String>> = futures_util::stream::iter(names.clone())
.map(|name| async move { newest_row(stream_ref, agent_ref, &name).await })
.buffered(LOOKUP_CONCURRENCY)
.collect()
.await;
@ -183,26 +190,52 @@ pub(crate) async fn list_subagents(
let entries = names
.into_iter()
.zip(lookups)
.map(|(name, lookup)| SubagentEntry {
.map(|(name, lookup)| {
let row = lookup.ok();
SubagentEntry {
name,
last_row_unix: lookup.ok(),
last_row_unix: row.as_ref().map(|r| r.unix),
model: row.as_ref().and_then(|r| r.model.clone()),
effort: row.and_then(|r| r.effort),
}
})
.collect();
Ok(Json(entries))
}
/// The publish time of a subagent's newest row, or why it could not be read.
/// A row that cannot be read, aged out between the listing and this lookup
/// say, is a missing time and not a failed list.
async fn last_row_unix(
/// What a subagent's newest row says: when it was published, and the model and
/// effort its headers name.
#[derive(Debug, PartialEq, Eq)]
struct NewestRow {
unix: i64,
model: Option<String>,
effort: Option<String>,
}
/// One header's value, or `None` when the row has none or an empty one.
fn header(headers: &async_nats::HeaderMap, name: &str) -> Option<String> {
headers
.get(name)
.map(ToString::to_string)
.filter(|v| !v.is_empty())
}
/// A subagent's newest row, or why it could not be read. A row that cannot be
/// read, aged out between the listing and this lookup say, is a missing time
/// and not a failed list.
async fn newest_row(
stream: &async_nats::jetstream::stream::Stream,
agent: &str,
subagent: &str,
) -> Result<i64, String> {
) -> Result<NewestRow, String> {
let subject = subagent_term::subject(agent, subagent);
let lookup = stream.get_last_raw_message_by_subject(&subject);
match tokio::time::timeout(LOOKUP_TIMEOUT, lookup).await {
Ok(Ok(row)) => Ok(row.time.unix_timestamp()),
Ok(Ok(row)) => Ok(NewestRow {
unix: row.time.unix_timestamp(),
model: header(&row.headers, subagent_term::HEADER_MODEL),
effort: header(&row.headers, subagent_term::HEADER_EFFORT),
}),
Ok(Err(e)) => Err(format!("{subagent}: {}", swarm_queue_client::chain(&e))),
Err(_) => Err(format!("{subagent}: no answer in {LOOKUP_TIMEOUT:?}")),
}
@ -291,7 +324,8 @@ fn read_failed(agent: &str, e: &dyn std::error::Error) -> problem_details::Probl
#[cfg(test)]
mod tests {
use swarm_wire::subagent_term::subagent_names;
use super::header;
use swarm_wire::subagent_term::{HEADER_EFFORT, HEADER_MODEL, subagent_names};
/// The list is what `STREAM.INFO` reports for the stream's subjects, by
/// name: a subagent that published once and one that published many rows
@ -307,4 +341,16 @@ mod tests {
["builder", "scout"]
);
}
/// The model and effort are read off the newest row's headers; a row
/// without them, or with an empty one, names nothing.
#[test]
fn the_model_and_effort_come_from_the_rows_headers() {
let mut headers = async_nats::HeaderMap::new();
headers.insert(HEADER_MODEL, "opus");
headers.insert(HEADER_EFFORT, "");
assert_eq!(header(&headers, HEADER_MODEL), Some("opus".to_owned()));
assert_eq!(header(&headers, HEADER_EFFORT), None);
assert_eq!(header(&async_nats::HeaderMap::new(), HEADER_MODEL), None);
}
}

View file

@ -32,6 +32,16 @@ const SUBAGENT_TOKEN: &str = "sub";
/// its last output.
pub const MAX_AGE: std::time::Duration = std::time::Duration::from_hours(24);
/// The message header carrying the model a subagent runs, on every row it
/// publishes. Absent when the subagent runs the harness's default model, which
/// is no value to name. The list endpoint reads it off the newest row, so a
/// subagent that runs for longer than [`MAX_AGE`] still shows it.
pub const HEADER_MODEL: &str = "Swarm-Model";
/// The message header carrying a subagent's reasoning effort, published with
/// [`HEADER_MODEL`].
pub const HEADER_EFFORT: &str = "Swarm-Effort";
/// The stream holding `agent`'s subagent rows: `term-sub-<agent>`.
#[must_use]
pub fn stream_name(agent: &str) -> String {