fix: verify gateway restart health after daemon restart

src/cli/daemon-cli/lifecycle.ts

@@ -1,13 +1,38 @@
+import { loadConfig, resolveGatewayPort } from "../../config/config.js";
 import { resolveGatewayService } from "../../daemon/service.js";
+import { defaultRuntime } from "../../runtime.js";
+import { theme } from "../../terminal/theme.js";
+import { formatCliCommand } from "../command-format.js";
 import {
   runServiceRestart,
   runServiceStart,
   runServiceStop,
   runServiceUninstall,
 } from "./lifecycle-core.js";
-import { renderGatewayServiceStartHints } from "./shared.js";
+import {
+  renderRestartDiagnostics,
+  terminateStaleGatewayPids,
+  waitForGatewayHealthyRestart,
+} from "./restart-health.js";
+import { parsePortFromArgs, renderGatewayServiceStartHints } from "./shared.js";
 import type { DaemonLifecycleOptions } from "./types.js";
 
+const POST_RESTART_HEALTH_ATTEMPTS = 8;
+const POST_RESTART_HEALTH_DELAY_MS = 450;
+
+async function resolveGatewayRestartPort() {
+  const service = resolveGatewayService();
+  const command = await service.readCommand(process.env).catch(() => null);
+  const serviceEnv = command?.environment ?? undefined;
+  const mergedEnv = {
+    ...(process.env as Record<string, string | undefined>),
+    ...(serviceEnv ?? undefined),
+  } as NodeJS.ProcessEnv;
+
+  const portFromArgs = parsePortFromArgs(command?.programArguments);
+  return portFromArgs ?? resolveGatewayPort(loadConfig(), mergedEnv);
+}
+
 export async function runDaemonUninstall(opts: DaemonLifecycleOptions = {}) {
   return await runServiceUninstall({
     serviceNoun: "Gateway",
@@ -41,11 +66,62 @@ export async function runDaemonStop(opts: DaemonLifecycleOptions = {}) {
  * Throws/exits on check or restart failures.
  */
 export async function runDaemonRestart(opts: DaemonLifecycleOptions = {}): Promise<boolean> {
+  const json = Boolean(opts.json);
+  const service = resolveGatewayService();
+  const restartPort = await resolveGatewayRestartPort().catch(() =>
+    resolveGatewayPort(loadConfig(), process.env),
+  );
+
   return await runServiceRestart({
     serviceNoun: "Gateway",
-    service: resolveGatewayService(),
+    service,
     renderStartHints: renderGatewayServiceStartHints,
     opts,
     checkTokenDrift: true,
+    postRestartCheck: async ({ warnings, fail, stdout }) => {
+      let health = await waitForGatewayHealthyRestart({
+        service,
+        port: restartPort,
+        attempts: POST_RESTART_HEALTH_ATTEMPTS,
+        delayMs: POST_RESTART_HEALTH_DELAY_MS,
+      });
+
+      if (!health.healthy && health.staleGatewayPids.length > 0) {
+        const staleMsg = `Found stale gateway process(es): ${health.staleGatewayPids.join(", ")}.`;
+        warnings.push(staleMsg);
+        if (!json) {
+          defaultRuntime.log(theme.warn(staleMsg));
+          defaultRuntime.log(theme.muted("Stopping stale process(es) and retrying restart..."));
+        }
+
+        await terminateStaleGatewayPids(health.staleGatewayPids);
+        await service.restart({ env: process.env, stdout });
+        health = await waitForGatewayHealthyRestart({
+          service,
+          port: restartPort,
+          attempts: POST_RESTART_HEALTH_ATTEMPTS,
+          delayMs: POST_RESTART_HEALTH_DELAY_MS,
+        });
+      }
+
+      if (health.healthy) {
+        return;
+      }
+
+      const diagnostics = renderRestartDiagnostics(health);
+      if (!json) {
+        defaultRuntime.log(theme.warn("Gateway did not become healthy after restart."));
+        for (const line of diagnostics) {
+          defaultRuntime.log(theme.muted(line));
+        }
+      } else {
+        warnings.push(...diagnostics);
+      }
+
+      fail("Gateway restart failed health checks.", [
+        formatCliCommand("openclaw gateway status --probe --deep"),
+        formatCliCommand("openclaw doctor"),
+      ]);
+    },
   });
 }