fix(maestro): snapshot /proc once per pass and check the whole group

The subtree search returned early whenever it found *any* group member,
so an in-tree stage that was busy or sleeping on a timer masked a
reparented member genuinely parked on the tty.

Checking every member instead would have meant a /proc scan per session
per tick whenever nothing was parked — the common case. ProcGroups
snapshots /proc once per activity pass and indexes by process group, so
the cost is one walk however many sessions the daemon owns, and the
subtree walk disappears: the index is a superset of it.

Also fixes the background-helper test, which had stopped discriminating.
It polled until awaiting_input was true and caught a moment before the
helper spawned, so it passed against a deepest-descendant walk too. It now
waits for the helper to exist first. Both regression tests were re-checked
by patching each old implementation back in.

Reported in review on #51.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

crates/maestro-daemon/src/proc.rs

@@ -1,8 +1,8 @@
 //! Best-effort `/proc` inspection for the activity watcher (NAP-340, Linux).
 //!
 //! Decides whether a session is **blocked waiting on the user**: ask the kernel
-//! which process group owns the terminal, then check that process is parked
-//! waiting on terminal input.
+//! which process group owns the terminal, then check whether any live member of
+//! that group is parked on the terminal.
 //!
 //! `/proc/<pid>/syscall` would say so exactly, but it's ptrace-gated and
 //! unreadable for our own sessions on a default kernel, so we combine three
@@ -16,97 +16,81 @@
 //! Every step degrades to `false` on a read failure, so a non-Linux host or a
 //! denied `/proc` read falls back to `Idle`.
 
-/// Whether the session on `tty` (e.g. `/dev/pts/3`) is blocked waiting on the
-/// user — the signal for `AwaitingInput`.
-pub fn awaiting_input(pane_pid: i32, tty: &str) -> bool {
-    let Some(pgid) = foreground_pgid(pane_pid) else {
-        return false;
-    };
-    foreground_members(pane_pid, pgid)
-        .into_iter()
-        .any(|pid| parked_on_tty(pid, tty))
+use std::collections::HashMap;
+
+/// One walk of `/proc`, shared by every session in an activity pass.
+///
+/// The foreground process group is the unit that matters, and its members are
+/// not reliably reachable from the pane: the group leader can exit while others
+/// run on, and a stage that forks a terminal-reading child and exits leaves
+/// that child reparented to init. Walking the pane's process tree misses both.
+///
+/// Snapshotting once per pass costs one `/proc` walk however many sessions the
+/// daemon owns — cheaper than the per-session `tmux list-panes` fork that
+/// already runs on the same tick.
+#[derive(Debug, Default)]
+pub struct ProcGroups {
+    stats: HashMap<i32, Stat>,
+    by_pgid: HashMap<i32, Vec<i32>>,
 }
 
-/// The process group the terminal considers foreground.
-///
-/// `tpgid` is the kernel's own answer, which is why this replaced walking
-/// `children` to the deepest descendant: that guessed the newest child was the
-/// foreground one, so an agent with a background helper looked idle.
-fn foreground_pgid(pane_pid: i32) -> Option<i32> {
-    let tpgid = read_stat(pane_pid)?.tpgid;
+impl ProcGroups {
+    pub fn snapshot() -> Self {
+        let mut stats = HashMap::new();
+        let mut by_pgid: HashMap<i32, Vec<i32>> = HashMap::new();
+        let Ok(entries) = std::fs::read_dir("/proc") else {
+            return Self::default();
+        };
+        for pid in entries
+            .flatten()
+            .filter_map(|e| e.file_name().to_str()?.parse::<i32>().ok())
+        {
+            // A process can exit mid-walk; skipping it is the correct answer.
+            if let Some(stat) = read_stat(pid) {
+                by_pgid.entry(stat.pgrp).or_default().push(pid);
+                stats.insert(pid, stat);
+            }
+        }
+        Self { stats, by_pgid }
+    }
+
+    /// The process group the terminal considers foreground, per `tpgid`.
+    fn foreground_pgid(&self, pane_pid: i32) -> Option<i32> {
+        let tpgid = self.stats.get(&pane_pid)?.tpgid;
         // -1 when the process has no controlling terminal.
         (tpgid > 0).then_some(tpgid)
     }
 
-/// Cap on processes examined per pass. A session's tree is a shell plus a
-/// handful of children; this only bounds a pathological one.
-const MAX_TREE_NODES: usize = 256;
+    fn members(&self, pgid: i32) -> &[i32] {
+        self.by_pgid.get(&pgid).map_or(&[], Vec::as_slice)
+    }
 
-/// Live members of process group `pgid`.
+    fn state(&self, pid: i32) -> Option<char> {
+        self.stats.get(&pid).map(|s| s.state)
+    }
+}
+
+/// Whether the session on `tty` (e.g. `/dev/pts/3`) is blocked waiting on the
+/// user — the signal for `AwaitingInput`.
 ///
-/// `pgid` names a *group*, not a process, so `/proc/<pgid>` may not exist: the
-/// leader can exit while another member still holds the terminal.
-///
-/// The pane's subtree answers this in the normal case and costs a handful of
-/// reads. A member can still end up outside it — a stage that forks a
-/// terminal-reading child and exits leaves that child reparented to init while
-/// it keeps the foreground `pgid` — so an empty subtree result falls back to
-/// scanning `/proc`. That scan only runs when the cheap path found nothing.
-fn foreground_members(pane_pid: i32, pgid: i32) -> Vec<i32> {
-    let in_subtree = subtree_members(pane_pid, pgid);
-    if !in_subtree.is_empty() {
-        return in_subtree;
-    }
-    scan_for_pgid(pgid)
-}
-
-/// Members of `pgid` among `pane_pid` and its descendants.
-fn subtree_members(pane_pid: i32, pgid: i32) -> Vec<i32> {
-    let mut found = Vec::new();
-    let mut queue = vec![pane_pid];
-    let mut visited = 0usize;
-    while let Some(pid) = queue.pop() {
-        visited += 1;
-        if visited > MAX_TREE_NODES {
-            break;
-        }
-        if read_stat(pid).is_some_and(|s| s.pgrp == pgid) {
-            found.push(pid);
-        }
-        queue.extend(children(pid));
-    }
-    found
-}
-
-/// Every live process in `pgid`, by walking `/proc`. The fallback for a member
-/// that has been reparented out of the pane's subtree.
-fn scan_for_pgid(pgid: i32) -> Vec<i32> {
-    let Ok(entries) = std::fs::read_dir("/proc") else {
-        return Vec::new();
+/// True when *any* member of the terminal's foreground group is parked on it: a
+/// group can hold a busy or timer-sleeping stage alongside the one actually
+/// reading, and either can be the reparented one.
+pub fn awaiting_input(pane_pid: i32, tty: &str, groups: &ProcGroups) -> bool {
+    let Some(pgid) = groups.foreground_pgid(pane_pid) else {
+        return false;
     };
-    entries
-        .flatten()
-        .filter_map(|e| e.file_name().to_str()?.parse::<i32>().ok())
-        .filter(|pid| read_stat(*pid).is_some_and(|s| s.pgrp == pgid))
-        .collect()
-}
-
-/// Direct children of `pid`, via `/proc/<pid>/task/<pid>/children`.
-fn children(pid: i32) -> Vec<i32> {
-    std::fs::read_to_string(format!("/proc/{pid}/task/{pid}/children"))
-        .map(|s| {
-            s.split_whitespace()
-                .filter_map(|x| x.parse().ok())
-                .collect()
-        })
-        .unwrap_or_default()
+    groups
+        .members(pgid)
+        .iter()
+        .any(|pid| parked_on_tty(*pid, tty, groups))
 }
 
 /// Whether `pid` is parked on `tty` rather than busy or sleeping on a timer.
-fn parked_on_tty(pid: i32, tty: &str) -> bool {
+fn parked_on_tty(pid: i32, tty: &str, groups: &ProcGroups) -> bool {
     // Interruptible sleep is where a blocking read parks; anything else
     // (running, uninterruptible) is busy, not waiting on us.
-    if read_stat(pid).map(|s| s.state) != Some('S') {
+    if groups.state(pid) != Some('S') {
         return false;
     }
     // A pure timer sleep (`sleep`, `usleep`) also sits in 'S' with the terminal
@@ -234,31 +218,46 @@ mod tests {
         assert!(!is_timer_wchan("0"));
     }
 
-    /// The fallback used when no member of the foreground group is in the
-    /// pane's subtree. Exercised against this process's own group, which is
-    /// guaranteed to have at least one live member.
+    /// The snapshot indexes by process group, not by parentage — that is the
+    /// whole point, since a group member need not be under the pane. Checked
+    /// against this process's own group, which certainly has a live member.
     #[test]
-    fn scan_for_pgid_finds_this_process() {
+    fn snapshot_indexes_this_process_under_its_group() {
         // SAFETY: argument-less query of the caller's own process group.
         let own = unsafe { libc::getpgid(0) };
         assert!(own > 0, "expected a real process group, got {own}");
 
-        let members = scan_for_pgid(own);
+        let groups = ProcGroups::snapshot();
         let me = std::process::id() as i32;
         assert!(
-            members.contains(&me),
-            "scan missed this process ({me}) in group {own}: {members:?}"
+            groups.members(own).contains(&me),
+            "snapshot missed this process ({me}) in group {own}: {:?}",
+            groups.members(own)
+        );
+        assert!(
+            groups.state(me).is_some(),
+            "a process in the snapshot must carry its state"
         );
     }
 
     #[test]
-    fn scan_for_pgid_is_empty_for_a_group_that_cannot_exist() {
-        assert!(scan_for_pgid(i32::MAX).is_empty());
+    fn snapshot_has_no_members_for_a_group_that_cannot_exist() {
+        assert!(ProcGroups::snapshot().members(i32::MAX).is_empty());
+    }
+
+    #[test]
+    fn a_pid_with_no_controlling_terminal_has_no_foreground_group() {
+        let groups = ProcGroups::snapshot();
+        assert_eq!(groups.foreground_pgid(i32::MAX), None);
     }
 
     #[test]
     fn awaiting_input_is_false_for_nonexistent_pid() {
-        // A pid that can't exist → every read fails → false (fall back to Idle).
-        assert!(!awaiting_input(i32::MAX, "/dev/pts/0"));
+        // A pid that can't exist → not in the snapshot → false (falls back to Idle).
+        assert!(!awaiting_input(
+            i32::MAX,
+            "/dev/pts/0",
+            &ProcGroups::snapshot()
+        ));
     }
 }

crates/maestro-daemon/src/session.rs

@@ -595,6 +595,11 @@ async fn run_activity_pass(inner: &Arc<Inner>) {
         return;
     }
 
+    // One `/proc` walk for the whole pass, rather than one per session. Reading
+    // it is pure filesystem work, so it goes on the blocking pool alongside the
+    // tmux calls below.
+    let groups = tmux_blocking(crate::proc::ProcGroups::snapshot).await;
+
     for id in ids {
         // The tmux pane lookup is a subprocess — keep it off the runtime worker
         // and outside the sessions lock.
@@ -611,7 +616,7 @@ async fn run_activity_pass(inner: &Arc<Inner>) {
             }
         };
         // `/proc` inspection is cheap, but still done off the async lock.
-        let awaiting = crate::proc::awaiting_input(pane.pane_pid, &pane.tty);
+        let awaiting = crate::proc::awaiting_input(pane.pane_pid, &pane.tty, &groups);
 
         let updated = {
             let mut sessions = inner.sessions.lock().await;

crates/maestro-daemon/tests/activity.rs

@@ -10,7 +10,7 @@ mod common;
 use std::time::Duration;
 
 use common::TmuxFixture;
-use maestro_daemon::proc::awaiting_input;
+use maestro_daemon::proc::{awaiting_input, ProcGroups};
 use tempfile::TempDir;
 use uuid::Uuid;
 
@@ -22,7 +22,8 @@ fn wait_for_awaiting(fx: &TmuxFixture, id: Uuid, want: bool, label: &str) {
     let mut last = None;
     while std::time::Instant::now() < deadline {
         if let Ok(Some(pane)) = fx.tmux.pane_info(id) {
-            let got = awaiting_input(pane.pane_pid, &pane.tty);
+            // Fresh each poll: the tree is still settling.
+            let got = awaiting_input(pane.pane_pid, &pane.tty, &ProcGroups::snapshot());
             if got == want {
                 return;
             }
@@ -70,6 +71,15 @@ fn a_foreground_reader_with_a_background_helper_is_awaiting_input() {
         )
         .expect("new_session");
 
+    // Wait for the helper itself. Asserting before it exists would pass
+    // against a deepest-descendant walk too, because until then the newest
+    // process in the tree *is* the one blocked on read.
+    let pane = fx.tmux.pane_info(id).expect("pane_info").expect("pane");
+    wait_until(
+        || subtree_has_comm(pane.pane_pid, "sleep"),
+        "the background helper to be spawned",
+    );
+
     wait_for_awaiting(
         &fx,
         id,
@@ -92,6 +102,39 @@ fn a_foreground_sleep_is_not_awaiting_input() {
     wait_for_awaiting(&fx, id, false, "a timer sleep is not user input");
 }
 
+/// Whether any process under `root` (inclusive) has this `comm`.
+fn subtree_has_comm(root: i32, comm: &str) -> bool {
+    let mut queue = vec![root];
+    while let Some(pid) = queue.pop() {
+        if std::fs::read_to_string(format!("/proc/{pid}/comm")).is_ok_and(|c| c.trim() == comm) {
+            return true;
+        }
+        queue.extend(
+            std::fs::read_to_string(format!("/proc/{pid}/task/{pid}/children"))
+                .map(|s| {
+                    s.split_whitespace()
+                        .filter_map(|x| x.parse::<i32>().ok())
+                        .collect::<Vec<_>>()
+                })
+                .unwrap_or_default(),
+        );
+    }
+    false
+}
+
+/// Block until `cond` holds, so a test asserts on the state it means to
+/// exercise rather than on whatever the tree looked like first.
+fn wait_until(cond: impl Fn() -> bool, what: &str) {
+    let deadline = std::time::Instant::now() + Duration::from_secs(5);
+    while std::time::Instant::now() < deadline {
+        if cond() {
+            return;
+        }
+        std::thread::sleep(Duration::from_millis(20));
+    }
+    panic!("timed out waiting for {what}");
+}
+
 /// Read `pgrp` and `tpgid` for a pid straight from `/proc`, so the test can
 /// wait for the exact condition it means to exercise rather than a proxy.
 fn pgrp_and_tpgid(pid: i32) -> Option<(i32, i32)> {
@@ -147,7 +190,7 @@ fn a_foreground_pipeline_whose_leader_exited_is_still_awaiting_input() {
     );
 
     assert!(
-        awaiting_input(pane.pane_pid, &pane.tty),
+        awaiting_input(pane.pane_pid, &pane.tty, &ProcGroups::snapshot()),
         "a surviving pipeline member still holds the tty"
     );
 }