[codex] degrade retryable gateway startup failures

gateway/run.py

@@ -3449,6 +3449,7 @@ async def start(self) -> bool:
         enabled_platform_count = 0
         startup_nonretryable_errors: list[str] = []
         startup_retryable_errors: list[str] = []
+        startup_degraded_reason: Optional[str] = None
 
         # Initialize and connect each configured platform
         for platform, platform_config in self.config.platforms.items():
@@ -3580,16 +3581,16 @@ async def start(self) -> bool:
                 return True
             if enabled_platform_count > 0:
                 if startup_retryable_errors:
-                    # At least one platform attempted a connection and failed —
-                    # this is a real startup error that should block the gateway.
+                    # All observed startup failures are retryable. Keep the
+                    # process alive so the reconnect watcher can recover instead
+                    # of relying on service-manager restart loops.
                     reason = "; ".join(startup_retryable_errors)
-                    logger.error("Gateway failed to connect any configured messaging platform: %s", reason)
-                    try:
-                        from gateway.status import write_runtime_status
-                        write_runtime_status(gateway_state="startup_failed", exit_reason=reason)
-                    except Exception:
-                        pass
-                    return False
+                    startup_degraded_reason = reason
+                    logger.warning(
+                        "Gateway failed to connect any messaging platform at startup; "
+                        "keeping process alive for reconnect watcher: %s",
+                        reason,
+                    )
                 # All enabled platforms had no adapter (missing library or credentials).
                 # In fleet deployments the same config.yaml is shared across nodes that
                 # may only have credentials for a subset of platforms.  Rather than
@@ -3609,7 +3610,10 @@ async def start(self) -> bool:
         self._wire_teams_pipeline_runtime()
 
         self._running = True
-        self._update_runtime_status("running")
+        if startup_degraded_reason and connected_count == 0 and self._failed_platforms:
+            self._update_runtime_status("degraded", startup_degraded_reason)
+        else:
+            self._update_runtime_status("running", None)
 
         # Emit gateway:startup hook
         hook_count = len(self.hooks.loaded_hooks)

tests/gateway/test_runner_startup_failures.py

@@ -64,7 +64,7 @@ async def get_chat_info(self, chat_id):
 
 
 @pytest.mark.asyncio
-async def test_runner_returns_failure_for_retryable_startup_errors(monkeypatch, tmp_path):
+async def test_runner_degrades_for_retryable_startup_errors(monkeypatch, tmp_path):
     monkeypatch.setenv("HERMES_HOME", str(tmp_path))
     config = GatewayConfig(
         platforms={
@@ -75,16 +75,21 @@ async def test_runner_returns_failure_for_retryable_startup_errors(monkeypatch,
     runner = GatewayRunner(config)
 
     monkeypatch.setattr(runner, "_create_adapter", lambda platform, platform_config: _RetryableFailureAdapter())
+    monkeypatch.setattr(runner.hooks, "discover_and_load", lambda: None)
+    monkeypatch.setattr(runner.hooks, "emit", AsyncMock())
 
     ok = await runner.start()
 
-    assert ok is False
+    assert ok is True
     assert runner.should_exit_cleanly is False
     state = read_runtime_status()
-    assert state["gateway_state"] == "startup_failed"
+    assert state["gateway_state"] == "degraded"
     assert "temporary DNS resolution failure" in state["exit_reason"]
     assert state["platforms"]["telegram"]["state"] == "retrying"
     assert state["platforms"]["telegram"]["error_code"] == "telegram_connect_error"
+    assert Platform.TELEGRAM in runner._failed_platforms
+
+    await runner.stop()
 
 
 @pytest.mark.asyncio