huskies: merge 852

server/src/agents/pool/pipeline/merge/runner.rs

@@ -10,23 +10,17 @@ use super::time::{
 };
 
 impl AgentPool {
-    /// Start the merge pipeline as a background task.
+    /// Sweep all Running merge jobs in the CRDT and delete any that were left
+    /// behind by a previous server instance.
     ///
-    /// Returns immediately so the MCP tool call doesn't time out (the full
-    /// pipeline — squash merge + quality gates — takes well over 60 seconds,
-    /// exceeding Claude Code's MCP tool-call timeout).
+    /// A job is considered stale when its recorded `server_start` timestamp is
+    /// older than the current server's boot time, or when the `error` field
+    /// cannot be decoded (legacy / malformed entries).
     ///
-    /// The mergemaster agent should poll [`get_merge_status`](Self::get_merge_status)
-    /// until the job reaches a terminal state.
-    pub fn start_merge_agent_work(
-        self: &Arc<Self>,
-        project_root: &Path,
-        story_id: &str,
-    ) -> Result<(), String> {
-        // Sweep stale Running entries left behind by dead processes before
-        // applying the double-start guard.  This handles the case where the
-        // server crashed mid-merge: the next attempt finds a Running entry
-        // whose owning process is gone and clears it automatically.
+    /// Called at the top of [`start_merge_agent_work`] to unblock retries, and
+    /// also by the periodic background reaper in the tick loop so stale entries
+    /// are cleaned up even when no new merge is triggered.
+    pub(crate) fn reap_stale_merge_jobs(&self) {
         if let Some(jobs) = crate::crdt_state::read_all_merge_jobs() {
             let current_boot = server_start_time();
             for job in jobs {
@@ -46,6 +40,24 @@ impl AgentPool {
                 }
             }
         }
+    }
+
+    /// Start the merge pipeline as a background task.
+    ///
+    /// Returns immediately so the MCP tool call doesn't time out (the full
+    /// pipeline — squash merge + quality gates — takes well over 60 seconds,
+    /// exceeding Claude Code's MCP tool-call timeout).
+    ///
+    /// The mergemaster agent should poll [`get_merge_status`](Self::get_merge_status)
+    /// until the job reaches a terminal state.
+    pub fn start_merge_agent_work(
+        self: &Arc<Self>,
+        project_root: &Path,
+        story_id: &str,
+    ) -> Result<(), String> {
+        // Sweep stale Running entries left behind by dead processes before
+        // applying the double-start guard.
+        self.reap_stale_merge_jobs();
 
         // Guard against double-starts; clear any completed/failed entry so the
         // caller can retry without needing to call a separate cleanup step.