subagents, ACP: stop with SIGINT only, so claude ends its own tool commands

claude's Bash tool starts every command in a new session, so a signal to
claude's process group never reaches it. claude stops those commands
itself when SIGINT makes it shut down. A SIGKILL leaves them running,
re-parented to init. Per the ruling on #4999, every stop is SIGINT and
nothing escalates.

- Subagent `interrupt` always sends SIGINT. The `force` argument is gone
  from the MCP tool and from `session::interrupt`. A caller that still
  sends `force: true` gets SIGINT; the argument is ignored.
- `interrupt` used to free the session name before it signalled, so a
  `continue` could start a second claude on the session while the first
  was still exiting. The `running` map now holds a three-state `Slot`
  (starting, running, stopping). An interrupted claude turn is
  `Stopping` until the turn task's own `finish_turn` clears it. `status`
  says "stopping", and `start`/`continue` on that name are refused.
  `interrupt` still returns at once. The ACP arm is unchanged: its name
  is freed when `session/cancel` is sent.
- ACP `Connection` drop sends its process group SIGINT, not SIGKILL, and
  the spawn no longer sets tokio's `kill_on_drop`, which SIGKILLed the
  leader after the `Drop` body. The test fixture's background grandchild
  resets SIGINT to default, since `&` starts it ignored.

Tests: a fake claude that records SIGINT, interrupted, fails with
`cancel(true)` (the old `force: true` path). An interrupted fake claude
that keeps running keeps its name claimed and `continue` is refused,
which fails when the slot is freed before the signal. A fake ACP agent
that records SIGINT, dropped, fails with SIGKILL in `Drop`, and also with
SIGINT in `Drop` but `kill_on_drop(true)` left in place.

Not changed: hive-claude's idle watchdog (still SIGKILL, separate
hive-claude change), the units' `KillSignal`, `machinectl kill`, and the
login helper.

hive-runtime/README.md

@@ -35,13 +35,15 @@ The crate depends on no hyperhive binary crate, so `hive-agent` and
 A turn is stopped with `session/cancel`: by the `Canceller` handle, or by the
 idle watchdog once no `session/update` has arrived for
 `Config::idle_timeout`. An agent that has not answered the prompt 10s later
-is killed, and the next turn respawns it. A watchdog stop fails the turn
-with `AcpError::IdleTimeout` (`IdleKilled` if the agent was killed). This is
+is sent SIGINT, and the next turn respawns it. A watchdog stop fails the turn
+with `AcpError::IdleTimeout` (`IdleKilled` if the agent was sent SIGINT). This is
 the only way a provider error the agent retries without reporting it, such as
 an HTTP 429, ends the turn.
 
-The agent leads its own process group. Killing it, and dropping the runtime,
-`SIGKILL`s that group, so the commands the agent started die with it.
+The agent leads its own process group. Stopping it, and dropping the runtime,
+sends that group SIGINT, so the commands the agent started get it too.
+Nothing escalates to SIGKILL: an agent that ignores SIGINT keeps running,
+and the next turn starts a new one beside it.
 
 ## ACP backend: compaction
 

hive-runtime/src/acp/mod.rs

@@ -39,7 +39,7 @@ const SETTLE_QUIET: Duration = Duration::from_millis(500);
 const SETTLE_MAX: Duration = Duration::from_secs(5);
 
 /// How long the agent has to end its turn after `session/cancel` before it is
-/// killed.
+/// sent SIGINT.
 const CANCEL_GRACE: Duration = Duration::from_secs(10);
 
 /// How long a `compact` command may send nothing before it is cancelled and
@@ -120,10 +120,12 @@ pub enum AcpError {
     EmptyEndTurn,
     #[error(
         "ACP agent sent nothing for {silent_secs}s and did not stop within {grace_secs}s of \
-         `session/cancel`, so it was killed"
+         `session/cancel`, so it was sent SIGINT"
     )]
     IdleKilled { silent_secs: u64, grace_secs: u64 },
-    #[error("ACP agent did not stop within {grace_secs}s of `session/cancel`, so it was killed")]
+    #[error(
+        "ACP agent did not stop within {grace_secs}s of `session/cancel`, so it was sent SIGINT"
+    )]
     CancelIgnored { grace_secs: u64 },
 }
 
@@ -141,8 +143,8 @@ struct CancelState {
 impl Canceller {
     /// Ask the turn in flight to stop. The agent is sent `session/cancel`,
     /// and the turn then ends normally with what it produced so far; an
-    /// agent that has not ended it within `CANCEL_GRACE` is killed, and the
-    /// turn fails with [`AcpError::CancelIgnored`]. Returns whether a turn
+    /// agent that has not ended it within `CANCEL_GRACE` is sent SIGINT, and
+    /// the turn fails with [`AcpError::CancelIgnored`]. Returns whether a turn
     /// was in flight; with none, this does nothing.
     #[must_use]
     pub fn cancel(&self) -> bool {
@@ -1062,9 +1064,14 @@ mod tests {
     const AGENT: &str = r#"
 n=0 prompt= model=m/think effort=low
 printf 'start\n' >> "$1.methods"
+# With TRAP_INT set: SIGINT creates <log>.int and exits 0.
+if [ -n "$TRAP_INT" ]; then
+  trap ': > "$1.int"; exit 0' INT
+fi
 # With GRANDCHILD set: a background `sleep`, its pid in <log>.grandchild.
+# `&` starts it with SIGINT ignored; `env` puts the default back.
 if [ -n "$GRANDCHILD" ]; then
-  sh -c 'echo $$ >"$1.grandchild"; exec sleep 300' sh "$1" >/dev/null 2>&1 &
+  sh -c 'echo $$ >"$1.grandchild"; exec env --default-signal=INT sleep 300' sh "$1" >/dev/null 2>&1 &
 fi
 update() {
   printf '{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"%s","update":%s}}\n' "$1" "$2"
@@ -1529,6 +1536,33 @@ done
         assert_gone(grandchild).await;
     }
 
+    /// SIGINT rather than SIGKILL: an agent only stops the commands it
+    /// started itself if it gets to shut down.
+    #[tokio::test]
+    async fn dropping_the_runtime_lets_its_agent_shut_down_on_sigint() {
+        let dir = tempfile::tempdir().unwrap();
+        let runtime = agent(
+            dir.path(),
+            "ok",
+            &[("TRAP_INT", "1")],
+            PercentPolicy::default(),
+        );
+        runtime
+            .run(&config(dir.path()), "one", &NoopSink)
+            .await
+            .unwrap();
+
+        drop(runtime);
+        let marker = dir.path().join("prompts.int");
+        for _ in 0..250 {
+            if marker.exists() {
+                break;
+            }
+            tokio::time::sleep(Duration::from_millis(20)).await;
+        }
+        assert!(marker.exists(), "the agent never got SIGINT");
+    }
+
     fn sent(session: &str, text: &str) -> (String, String) {
         (session.to_owned(), text.to_owned())
     }

hive-runtime/src/acp/rpc.rs

@@ -31,9 +31,10 @@ type Reply = Result<Value, AcpError>;
 type Pending = Arc<Mutex<HashMap<u64, (&'static str, oneshot::Sender<Reply>)>>>;
 
 /// The agent runs as the leader of its own process group; dropping the
-/// connection `SIGKILL`s that group, so the commands the agent started die
-/// with it. A descendant that left the group (`setsid`, `setpgid`) is not
-/// reached.
+/// connection sends that group SIGINT, so the agent can stop the commands
+/// it started, and they get the signal too. Nothing escalates: an agent
+/// that ignores SIGINT keeps running. A descendant that left the group
+/// (`setsid`, `setpgid`) is not reached.
 pub(super) struct Connection {
     child: Child,
     /// The agent's pid, which is its group's id. Read at spawn: a reaped
@@ -66,7 +67,6 @@ impl Connection {
             .stdout(Stdio::piped())
             .stderr(Stdio::piped())
             .process_group(0)
-            .kill_on_drop(true)
             .spawn()
             .map_err(|source| AcpError::Spawn {
                 program: command.command.clone(),
@@ -182,7 +182,7 @@ impl Drop for Connection {
             // SAFETY: plain `kill(2)`; `-pgid` targets the agent's own
             // process group, which `process_group(0)` made it lead.
             unsafe {
-                libc::kill(-pgid, libc::SIGKILL);
+                libc::kill(-pgid, libc::SIGINT);
             }
         }
     }