This looks like a regression of the Windows gateway probe / connect-timeout class of failures that was previously marked fixed around 2026.3.22.

Related locked/closed issues:

• [Bug]: local gateway CLI handshake fails (probe timeout / gateway call closed 1000) even though gateway is running and WS challenge is reachable #45560: gateway probe timeout while gateway healthy
• [Bug]: Local CLI operator commands intermittently fail with "connect challenge timeout" / "handshake timeout" (code 1000/1008) under gateway load #49726: connect challenge / handshake timeout under gateway load

Environment:

• Windows: Microsoft Windows NT 10.0.26200.0
• Install: npm global
• OpenClaw latest tested: 2026.4.26
• Stable workaround: 2026.4.23 (a979721)
• Node runtime used by gateway: v22.22.2
• Codex CLI: 0.125.0
• Gateway: local, loopback, token auth, port 18789

Symptoms on 2026.4.26:

• Gateway process starts and listens on 127.0.0.1:18789.
• Logs show http server listening / gateway ready.
• openclaw gateway probe --timeout 60000 reports loopback connect timeout / unreachable.
• Gateway logs show handshake-timeout with lastFrameMethod:"connect".
• This persisted with Telegram disabled and with minimal/near-zero plugins.
• In contrast, Control UI/webchat and runtime RPC paths could still work, so gateway probe became an unreliable health source.
• A watchdog relying on gateway probe repeatedly restarted an otherwise usable gateway.

Additional Windows-specific finding:

• Bundled provider extensions anthropic and lmstudio attempted to load even though they were disabled in config.
• They failed with ERR_PACKAGE_PATH_NOT_EXPORTED from their bundled @mariozechner/pi-ai dependency.
• Those load failures appeared to block startup/handshake long enough to trigger more WS handshake timeouts.
• Quarantining those two broken bundled extension directories plus using longer OPENCLAW_HANDSHAKE_TIMEOUT_MS / OPENCLAW_CONNECT_CHALLENGE_TIMEOUT_MS helped stabilize the local install.

Working workaround:

• Downgrade/pin OpenClaw to 2026.4.23.
• Use channels status --deep / channels status --probe as the watchdog health check instead of gateway probe, because gateway probe was a false negative.
• Keep gateway handshake timeout env vars high on Windows.
• Keep the broken bundled provider extensions out of the load path until their packaging/export issue is fixed.

Why I think this should be treated as a regression:

• The previously fixed handshake/probe failure is reproducible again on a newer Windows build.
• The strongest signal is the same contradiction described in prior issues: gateway/channel runtime healthy and Telegram working, while gateway probe reports timeout/unreachable.
• There may be two interacting issues: the CLI/probe WS handshake race, plus disabled/broken provider extension loading causing event-loop stalls during gateway/connect handling.

Sanitized key signatures:

• gateway ready
• gateway probe: Local loopback ws://127.0.0.1:18789 -> Connect: failed - timeout
• gateway log: handshake timeout ... lastFrameMethod=connect
• provider load errors: anthropic/lmstudio failed to load ... ERR_PACKAGE_PATH_NOT_EXPORTED ... @mariozechner/pi-ai

I can provide more sanitized logs if useful.