@volter/supercode-orchestrator 0.4.104 → 0.4.105

package/board/dispatch.mjs

@@ -176,6 +176,16 @@ export function lanesOf(profiles = {}) {
 
 function supervise(board, fleet, settings, report, entry, effects) {
   const rows = board.db.prepare(entry.workflow.run ? "SELECT id FROM tasks WHERE current_run_id IS NOT NULL" : "SELECT id FROM tasks WHERE status = 'running'").all();
+  // The catalog can lag a pane. Read the actual machines once, before any board mutation.
+  const observed = new Map();
+  for (const { id } of rows) {
+    const s = currentRun(board.db, id)?.metadata?.supercode;
+    if (!s?.pane || observed.has(s.machine)) continue;
+    const r = sc(['teams', 'panes', 'ls', '--machine', s.machine ?? localMachine()], { timeout: 60_000 });
+    const list = r.code === 0 ? json(r.stdout.trim(), null) : null;
+    observed.set(s.machine, Array.isArray(list) ? new Set(list.map((p) => p.pane)) : null);
+    if (!Array.isArray(list)) report.failed.push([id, `pane liveness unavailable; retaining run: ${(r.stderr || r.stdout).trim().slice(-200)}`]);
+  }
   for (const { id } of rows) {
     tx(board, (db) => {
       const task = getTask(db, id);
@@ -197,7 +207,9 @@ function supervise(board, fleet, settings, report, entry, effects) {
       };
       const limits = limitsOf(db, board, entry, task, session);
       // Hermes's orphan reconciliation: a running card whose claim bookkeeping is broken, with nothing live behind it
-      const live = session?.surface === 'headless' ? pidAlive(session.pid) : session?.session_id ? (fleet.ok ? sessionAlive(fleet, session) : true) : false;
+      const paneReading = session?.pane ? observed.get(session.machine) : undefined;
+      const paneLive = paneReading instanceof Set ? paneReading.has(session.pane) : null;
+      const live = session?.surface === 'headless' ? pidAlive(session.pid) : session?.pane ? paneLive !== false : session?.session_id ? (fleet.ok ? sessionAlive(fleet, session) : true) : false;
       // (whether it is reconciled is the workflow's: its `orphaned` transitions and their guard)
       if ((!task.claim_lock || !task.claim_expires) && !live) {
         observe('orphaned', { reason: 'orphaned_running', claim_lock: task.claim_lock, claim_expires: task.claim_expires, error: 'orphaned running card (broken claim bookkeeping)' }, 'reclaimed');
@@ -226,20 +238,21 @@ function supervise(board, fleet, settings, report, entry, effects) {
         return;
       }
       if (session?.session_id) {
-        if (!fleet.ok) return; // no reading, no verdict
-        const alive = sessionAlive(fleet, session);
-        const machineUp = fleet.online.has(session.machine);
+        if (session.pane && paneLive === null) return; // no machine reading, no loss verdict
+        if (!session.pane && !fleet.ok) return;
+        const alive = session.pane ? paneLive : sessionAlive(fleet, session);
+        const machineUp = paneReading instanceof Set || fleet.online.has(session.machine);
         const young = t - Number(session.launched_at ?? 0) < Number(limits.grace ?? settings.grace);
         const missing = !alive && machineUp && !young;
         const next = { ...session };
         // a machine that is off the fleet: how long, for the workflow's machine_offline (the claim is kept meanwhile)
-        if (machineUp) delete next.offline_since; else next.offline_since ??= t;
+        if (machineUp) next.offline_since = null; else next.offline_since ??= t;
         if (!machineUp) {
           setRunSession(db, runRow.id, next);
           observe('machine_offline', { machine: session.machine, seconds: t - next.offline_since }, null);
           if (getTask(db, id)?.current_run_id !== runRow.id) return; // the workflow moved it
         }
-        if (alive) { next.seen_at ??= t; delete next.missing_since; }
+        if (alive) { next.seen_at ??= t; next.missing_since = null; }
         if (missing) next.missing_since ??= t;
         if (missing && t - next.missing_since >= Number(limits.grace ?? settings.grace)) {
           // lost: gone from an online machine for the whole grace. A resume that never came up is not resumable.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@volter/supercode-orchestrator",
-  "version": "0.4.104",
+  "version": "0.4.105",
   "type": "module",
   "description": "The orchestrator runtime over the Volter Harness ontology: one typed operational model whose folder is its serialization, read and written through the harness orchestration doors (docs/ORCHESTRATOR-IR.md)",
   "exports": {