mtech-ai-local-runner 0.1.40 → 0.1.41

package/README.md

@@ -1,3 +1,11 @@
+## 0.1.41: confirm a cleanup the guardian finished before a late terminate
+
+When Codex exits by itself, the owned guardian confirms the cleanup and exits. A busy daemon could then write its
+terminate request to the closed control socket before reading `cleaned`. That write error turned a confirmed cleanup
+into an unconfirmed one and suspended the job. A control socket error alone no longer does that: cleanup is
+confirmed by `cleaned` or by the guardian's exit code 0, which it uses only after sending `cleaned`. Timeouts,
+liveness errors, malformed messages and `unconfirmed` reports still fail closed.
+
 ## 0.1.40: say why the owned guardian could not confirm cleanup
 
 When the owned guardian reports an unconfirmed cleanup, the `executor_cleanup` diagnostic now carries fixed

package/dist/src/process/owned-process-guardian.js

@@ -78,14 +78,18 @@ export function spawnOwnedGuardian(command, args, options) {
     // Hosts can attach observers after spawn returns without unhandled rejection.
     void started.catch(() => { });
     let confirmed = false, spawnFailed = false, protocolFailed = false, buffer = '';
-    let terminateWritten = false, hostTimeout = false, detail;
+    let terminateWritten = false, hostTimeout = false, controlFailed = false, detail;
     let outcome;
     let closePromise;
     const closed = new Promise(resolve => child.once('close', () => { liveness.destroy(); resolve(); }));
+    // A guardian that has confirmed and exited closes its control socket; a terminate written late by a
+    // busy host then fails, possibly before "cleaned" is read. That write error alone is not a protocol
+    // failure: the guardian exits 0 only after it sent "cleaned" (or after a spawn failure owned nothing).
     control.on('error', () => { if (!confirmed && !spawnFailed)
-        protocolFailed = true; });
+        controlFailed = true; });
     liveness.on('error', () => { if (!confirmed && !spawnFailed)
         protocolFailed = true; });
+    const cleanedByExit = () => controlFailed && !protocolFailed && child.exitCode === 0 && child.signalCode === null;
     control.on('data', (chunk) => {
         buffer += chunk;
         if (buffer.length > 8192) {
@@ -140,7 +144,7 @@ export function spawnOwnedGuardian(command, args, options) {
             this.signal(immediate);
             closePromise = new Promise((resolve, reject) => {
                 const timer = setTimeout(() => { protocolFailed = true; hostTimeout = true; liveness.destroy(); reject(new Error('Provider process group cleanup could not be confirmed')); }, 5500);
-                void closed.then(() => { clearTimeout(timer); if (protocolFailed || !confirmed)
+                void closed.then(() => { clearTimeout(timer); if (protocolFailed || !(confirmed || cleanedByExit()))
                     reject(new Error('Provider process group cleanup could not be confirmed'));
                 else
                     resolve(); });
@@ -149,15 +153,17 @@ export function spawnOwnedGuardian(command, args, options) {
         },
         /** Fixed cleanup observations after an unconfirmed cleanup; empty once cleanup is confirmed. */
         cleanupObservations() {
-            if (confirmed && !protocolFailed)
+            if ((confirmed || cleanedByExit()) && !protocolFailed)
                 return [];
-            if (!protocolFailed && !hostTimeout)
+            if (!protocolFailed && !hostTimeout && !controlFailed)
                 return [];
             const observed = [terminateWritten ? 'cleanup_guardian_terminate_written' : 'cleanup_guardian_terminate_not_written'];
             if (hostTimeout)
                 observed.push('cleanup_guardian_host_timeout');
+            if (controlFailed)
+                observed.push('cleanup_guardian_control_failed');
             if (!detail) {
-                if (!hostTimeout)
+                if (!hostTimeout && protocolFailed)
                     observed.push('cleanup_guardian_protocol_failed');
                 return observed;
             }

package/dist/src/safety/execution-failure.js

@@ -20,7 +20,7 @@ export const STARTUP_OBSERVATIONS = Object.freeze([
     'cleanup_exec_tools_revoke_failed', 'cleanup_job_tools_revoke_failed', 'cleanup_stdin_end_failed', 'cleanup_process_terminate_failed',
     'cleanup_tool_server_close_failed', 'cleanup_timeout',
     // The owned guardian's own account of an unconfirmed cleanup (fixed words, no PIDs or paths).
-    'cleanup_guardian_terminate_written', 'cleanup_guardian_terminate_not_written', 'cleanup_guardian_host_timeout', 'cleanup_guardian_protocol_failed',
+    'cleanup_guardian_terminate_written', 'cleanup_guardian_terminate_not_written', 'cleanup_guardian_host_timeout', 'cleanup_guardian_control_failed', 'cleanup_guardian_protocol_failed',
     'cleanup_guardian_trigger_terminate', 'cleanup_guardian_trigger_root_exit', 'cleanup_guardian_trigger_lost', 'cleanup_guardian_trigger_output_error', 'cleanup_guardian_trigger_spawn_error',
     'cleanup_guardian_term_sent', 'cleanup_guardian_term_eperm', 'cleanup_guardian_term_esrch', 'cleanup_guardian_term_not_needed',
     'cleanup_guardian_kill_sent', 'cleanup_guardian_kill_eperm', 'cleanup_guardian_kill_esrch', 'cleanup_guardian_kill_not_needed',

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mtech-ai-local-runner",
-  "version": "0.1.40",
+  "version": "0.1.41",
   "private": false,
   "license": "UNLICENSED",
   "bin": {