huskies: merge 1170 bug Full tokio runtime stall after unblock → merge auto-assign

server/src/agents/pool/auto_assign/merge.rs

@@ -65,27 +65,43 @@ impl AgentPool {
             // AC6: Detect empty-diff stories before starting the merge pipeline.
             // If the worktree has no commits on the feature branch, block the
             // story immediately via the state machine — no merge job needed.
-            if let Some(wt_path) = worktree::find_worktree_path(project_root, story_id)
-                && !crate::agents::gates::worktree_has_committed_work(&wt_path)
-            {
-                let empty_diff_reason = "Feature branch has no code changes — the coder agent \
-                     did not produce any commits.";
-                slog_warn!(
-                    "[auto-assign] Story '{story_id}' in 4_merge/ has no commits \
-                     on feature branch. Blocking via state machine."
-                );
-                if let Err(e) =
-                    crate::agents::lifecycle::transition_to_blocked(story_id, empty_diff_reason)
-                {
-                    slog_error!("[auto-assign] Failed to transition '{story_id}' to Blocked: {e}");
+            //
+            // Bug 1170: worktree_has_committed_work shells out to `git log`
+            // synchronously. assign_merge_stage runs on the shared tokio
+            // runtime (it's invoked reactively on every CRDT transition, incl.
+            // unblock), so calling it inline here blocked a runtime worker
+            // thread for the duration of the git subprocess — with a story
+            // whose worktree/agent had crashed, that call could hang
+            // indefinitely and stall /health and the liveness heartbeat along
+            // with it. Run it on the blocking-thread pool instead.
+            if let Some(wt_path) = worktree::find_worktree_path(project_root, story_id) {
+                let has_commits = tokio::task::spawn_blocking(move || {
+                    crate::agents::gates::worktree_has_committed_work(&wt_path)
+                })
+                .await
+                .unwrap_or(false);
+                if !has_commits {
+                    let empty_diff_reason = "Feature branch has no code changes — the coder agent \
+                         did not produce any commits.";
+                    slog_warn!(
+                        "[auto-assign] Story '{story_id}' in 4_merge/ has no commits \
+                         on feature branch. Blocking via state machine."
+                    );
+                    if let Err(e) =
+                        crate::agents::lifecycle::transition_to_blocked(story_id, empty_diff_reason)
+                    {
+                        slog_error!(
+                            "[auto-assign] Failed to transition '{story_id}' to Blocked: {e}"
+                        );
+                    }
+                    let _ = self
+                        .watcher_tx
+                        .send(crate::io::watcher::WatcherEvent::StoryBlocked {
+                            story_id: story_id.to_string(),
+                            reason: empty_diff_reason.to_string(),
+                        });
+                    continue;
                 }
-                let _ = self
-                    .watcher_tx
-                    .send(crate::io::watcher::WatcherEvent::StoryBlocked {
-                        story_id: story_id.to_string(),
-                        reason: empty_diff_reason.to_string(),
-                    });
-                continue;
             }
 
             // Skip if a merge job is already running for this story (e.g. triggered
@@ -111,3 +127,89 @@ impl AgentPool {
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::super::super::AgentPool;
+    use crate::config::ProjectConfig;
+    use std::sync::Arc;
+    use std::sync::atomic::{AtomicU64, Ordering};
+
+    /// Bug 1170 regression: `assign_merge_stage` used to call
+    /// `worktree_has_committed_work` (which shells out to `git`) directly on
+    /// the async runtime. For a story with a crashed/unassignable agent
+    /// sitting in `4_merge/`, that synchronous subprocess call had no yield
+    /// point, so on a runtime with few worker threads it starved every other
+    /// task — including the liveness heartbeat and `/health` — for the whole
+    /// scan. After wrapping the call in `spawn_blocking`, the executor stays
+    /// free to interleave other work while the git subprocess runs
+    /// off-runtime.
+    ///
+    /// This reproduces the unblock → merge-auto-assign path: a story sits in
+    /// `4_merge/` with no active agent entry (the crashed/unassignable case)
+    /// and a worktree directory that isn't a real git repo, forcing every
+    /// `git` invocation in the scan to fail — but only after paying the
+    /// process fork/exec cost, which is what stalls a non-yielding runtime.
+    #[tokio::test(flavor = "multi_thread", worker_threads = 1)]
+    async fn assign_merge_stage_does_not_stall_liveness_heartbeat() {
+        crate::db::ensure_content_store();
+
+        let tmp = tempfile::tempdir().unwrap();
+        let root = tmp.path().to_path_buf();
+        std::fs::create_dir_all(root.join(".huskies")).unwrap();
+        std::fs::write(root.join(".huskies/project.toml"), "").unwrap();
+
+        let worktrees_dir = root.join(".huskies/worktrees");
+        std::fs::create_dir_all(&worktrees_dir).unwrap();
+
+        // Simulate several stories stuck in 4_merge/ with a crashed/unassignable
+        // agent: each has a worktree directory (so find_worktree_path succeeds
+        // and the git-shelling check runs) but is not a real git repo and has
+        // no active agent entry in the pool.
+        const STORY_COUNT: usize = 25;
+        for i in 0..STORY_COUNT {
+            let story_id = format!("11700_merge_{i:03}");
+            std::fs::create_dir_all(worktrees_dir.join(&story_id)).unwrap();
+            crate::db::write_item_with_content(
+                &story_id,
+                "4_merge",
+                "---\nname: Crashed Merge\n---\n",
+                crate::db::ItemMeta::named("Crashed Merge"),
+            );
+        }
+
+        let pool = AgentPool::new_test(3200);
+        let config = ProjectConfig::load(&root).unwrap_or_default();
+
+        // Stand in for the liveness heartbeat (tick_loop.rs's
+        // spawn_liveness_tick) and /health polling: a tight-interval task
+        // racing the merge scan on the single-worker-thread runtime.
+        let ticks = Arc::new(AtomicU64::new(0));
+        let ticks_clone = Arc::clone(&ticks);
+        let heartbeat = tokio::spawn(async move {
+            loop {
+                tokio::time::sleep(std::time::Duration::from_millis(2)).await;
+                ticks_clone.fetch_add(1, Ordering::SeqCst);
+            }
+        });
+
+        let start = std::time::Instant::now();
+        pool.assign_merge_stage(&root, &config).await;
+        let elapsed = start.elapsed();
+
+        heartbeat.abort();
+        let observed_ticks = ticks.load(Ordering::SeqCst);
+
+        // With a 2ms heartbeat cadence, an unstalled runtime should have
+        // fired roughly elapsed/2ms ticks. Require at least a quarter of that
+        // as a generous floor — a stalled runtime (pre-fix) produces ~0 ticks
+        // because the single worker thread never yields during the scan.
+        let expected_min_ticks = (elapsed.as_millis() / 2 / 4) as u64;
+        assert!(
+            observed_ticks >= expected_min_ticks,
+            "liveness heartbeat stalled during assign_merge_stage: {observed_ticks} tick(s) \
+             over {elapsed:?} (expected at least ~{expected_min_ticks}); the merge scan likely \
+             blocked the tokio runtime instead of yielding via spawn_blocking"
+        );
+    }
+}