agent-relay-runner 0.129.2 → 0.129.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-relay-runner",
-  "version": "0.129.2",
+  "version": "0.129.3",
   "description": "Unified provider lifecycle runner for Agent Relay",
   "type": "module",
   "bin": {

package/plugins/claude/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "agent-relay-runner",
   "description": "Thin Agent Relay runner bridge for Claude Code",
-  "version": "0.129.2",
+  "version": "0.129.3",
   "agentRelayContracts": {
     "providerPluginProtocol": 1
   }

package/src/runner-core.ts

@@ -1053,8 +1053,10 @@ export class AgentRunner {
         return;
       }
 
+      let restartResumeId: string | undefined;
       if (this.shouldStopUnexpectedProviderExit(diagnostics)) {
-        const hasResumeId = providerExitResumeId(diagnostics) !== undefined;
+        const resumeId = providerExitResumeId(diagnostics);
+        const hasResumeId = resumeId !== undefined;
         if (!hasResumeId) {
           const recoveredStatus = await this.recoveredStatusAfterUndeterminedProviderExit();
           if (recoveredStatus) {
@@ -1081,6 +1083,17 @@ export class AgentRunner {
         }
         if (!hasResumeId && this.shouldRestartRecoverableProviderExit(diagnostics)) {
           logger.warn("lifecycle", `${this.options.provider} exit signal is early/recoverable; allowing fresh restart`);
+        } else if (resumeId && this.shouldResumeWithContextProviderExit(diagnostics, recent.length)) {
+          // #1258 (L2): a long-lived agent killed via tmux-session-ended (external SIGTERM, reboot,
+          // OOM, admin cleanup) used to be stranded permanently offline even though the provider CLI printed
+          // a resumable session id. Now, when Layer 3 captured a valid resume id, auto-resume WITH
+          // context instead of surfacing manual recovery. Crash-loop protection reuses the existing
+          // unexpected-exit window: `recent` is capped at MAX_RAPID_UNEXPECTED_EXITS within
+          // UNEXPECTED_EXIT_WINDOW_MS, and the fall-through path below applies backoff + a hard
+          // give-up — so a genuinely broken agent stops resurrecting once the window fills, while a
+          // resume that survives past the window resets the counter (no permanent lockout).
+          restartResumeId = resumeId;
+          logger.warn("lifecycle", `${this.options.provider} exited via tmux; auto-resuming with captured context (resume id ${resumeId})`);
         } else {
           logger.warn("lifecycle", `${this.options.provider} exited; leaving agent offline for manual recovery`);
           this.publishRunnerTimelineEvent({
@@ -1141,18 +1154,21 @@ export class AgentRunner {
       }
 
       const delayMs = Math.min(10_000, Math.max(500, 500 * recent.length));
-      logger.warn("lifecycle", `provider session exited unexpectedly after ${Math.round(runtimeMs / 1000)}s; restarting in ${delayMs}ms`);
+      const resuming = restartResumeId !== undefined;
+      logger.warn("lifecycle", `provider session exited unexpectedly after ${Math.round(runtimeMs / 1000)}s; ${resuming ? "resuming with context" : "restarting"} in ${delayMs}ms`);
       this.publishRunnerTimelineEvent({
         status: "provider.restart_decision",
         id: `provider-restart-decision-${this.providerSessionId}-${now}`,
         timestamp: Date.now(),
         title: "Provider restart scheduled",
-        body: `runner will start a fresh ${this.options.provider} provider in ${delayMs}ms`,
+        body: resuming
+          ? `runner will resume ${this.options.provider} with captured context in ${delayMs}ms`
+          : `runner will start a fresh ${this.options.provider} provider in ${delayMs}ms`,
         icon: "ti-refresh",
         metadata: {
           eventType: "provider.restart_decision",
-          decision: "restart-fresh",
-          reason: "unexpected-headless-terminal-exit",
+          decision: resuming ? "resume-with-context" : "restart-fresh",
+          reason: resuming ? "tmux-exit-auto-resume" : "unexpected-headless-terminal-exit",
           delayMs,
           rapidExitCount: recent.length,
           rapidExitWindowMs: UNEXPECTED_EXIT_WINDOW_MS,
@@ -1162,7 +1178,7 @@ export class AgentRunner {
       await Bun.sleep(delayMs);
       if (this.stopped || this.exitCommandInProgress) return;
       try {
-        await this.restartProvider();
+        await this.restartProvider(restartResumeId ? { resumeId: restartResumeId } : undefined);
         this.publishStatus();
         this.scheduleDrain();
       } catch (error) {
@@ -1186,6 +1202,15 @@ export class AgentRunner {
       && diagnostics.runtimeMs < RAPID_EXIT_MS;
   }
 
+  // #1258 (L2): auto-resume-with-context gate for a tmux-session-ended death at any point in an
+  // agent's life (not just the <30s early-flake window). Requires a captured resume id (the caller
+  // only reaches this when one exists) and bounds crash loops via the shared unexpected-exit window:
+  // once `recent` exceeds MAX_RAPID_UNEXPECTED_EXITS the agent is left offline for manual recovery.
+  private shouldResumeWithContextProviderExit(diagnostics: Record<string, unknown>, recentUnexpectedExits: number): boolean {
+    return diagnostics.exitSource === "tmux-session-ended"
+      && recentUnexpectedExits <= MAX_RAPID_UNEXPECTED_EXITS;
+  }
+
   private async recoveredStatusAfterUndeterminedProviderExit(): Promise<"busy" | "idle" | undefined> {
     if (!this.process || !this.options.adapter.probeActivity) return undefined;
     const deadline = Date.now() + UNDETERMINED_PROVIDER_EXIT_CONFIRM_MS;

package/src/runner-helpers.ts

@@ -168,6 +168,17 @@ export function runnerShouldRestartUnexpectedProviderExit(
     && input.hasTerminalSession;
 }
 
+// #1258 (L3): claude prints its resume banner ANSI-dimmed — the raw log bytes are
+// `\x1b[2mclaude --resume <uuid>\x1b[22m`, so `claude` is immediately preceded by the SGR
+// escape's trailing `m`. A leading `\b` in the extraction regex finds no word boundary
+// (`m`→`c` are both word chars) and yields zero matches, stranding the resume id as null.
+// Strip CSI escape sequences before matching so the banner reads as plain text.
+const ANSI_ESCAPE_RE = /\x1b\[[0-9;?]*[ -/]*[@-~]/g;
+
+function stripAnsi(text: string): string {
+  return text.replace(ANSI_ESCAPE_RE, "");
+}
+
 function resumeLogExtractionRe(provider: string): RegExp | undefined {
   const manifest = getManifest(provider);
   const resume = manifest?.resume as { supported: boolean; idPattern?: string } | undefined;
@@ -178,8 +189,9 @@ function resumeLogExtractionRe(provider: string): RegExp | undefined {
 export function latestResumeIdFromText(provider: string, text: string): string | undefined {
   const re = resumeLogExtractionRe(provider);
   if (!re) return undefined;
+  const cleaned = stripAnsi(text);
   let latest: string | undefined;
-  for (let match = re.exec(text); match; match = re.exec(text)) {
+  for (let match = re.exec(cleaned); match; match = re.exec(cleaned)) {
     latest = match[1];
   }
   return latest;