openclaw - 💡(How to fix) Fix [Bug]: Windows 11 agent --local hangs on 2026.4.29 while capability model run succeeds [2 comments, 3 participants]

Summary

On native Windows, openclaw agent --local hangs past a 10-minute shell budget on OpenClaw 2026.4.29 (a448042) even when the same profile, auth, provider, and models work through capability model run --local in a few seconds. This looks below the model/auth layer, in the native Windows embedded agent runtime or command lane.

Environment

• OS: Windows 11 x64
• OpenClaw: 2026.4.29 (a448042)
• Node: 24.15.0 from official Windows x64 zip, invoked directly
• Profile: isolated lab profile imladri-lab
• OPENCLAW_DISABLE_BONJOUR=1
• Auth: valid persistent OpenAI auth profile openai:manual
• Default model after auth persistence: openai/gpt-4o-mini
• No channel configured; this is not WhatsApp-specific

What works

All of these succeeded with the same profile/auth path:

openclaw --profile imladri-lab capability model run --local --model openai/gpt-5.4-nano --prompt "Return exactly: OPENCLAW_GPT54_NANO_OK" --json
openclaw --profile imladri-lab capability model run --local --model openai/gpt-4.1-mini --prompt "Return exactly: OPENCLAW_GPT41_MINI_OK" --json
openclaw --profile imladri-lab capability model run --local --model openai/gpt-4o-mini --prompt "Return exactly: OPENCLAW_GPT4O_MINI_OK" --json

After persisting auth and setting the default model, this also succeeded without a process-scoped API key:

openclaw --profile imladri-lab models set openai/gpt-4o-mini
openclaw --profile imladri-lab capability model run --local --prompt "Return exactly: OPENCLAW_DEFAULT_MODEL_OK" --json

The response reported provider openai, model gpt-4o-mini, capability model.run, transport local, and the expected text.

What hangs

These full agent runs authenticated but did not complete:

openclaw --profile imladri-lab agent --local --session-id imladri-gpt54-nano-smoke --model openai/gpt-5.4-nano --thinking off --message "Return exactly: OPENCLAW_AGENT_OK" --json --timeout 600
openclaw --profile imladri-lab agent --local --session-id imladri-gpt41-mini-smoke --model openai/gpt-4.1-mini --thinking off --message "Return exactly: OPENCLAW_AGENT_OK" --json --timeout 180
openclaw --profile imladri-lab agent --local --session-id imladri-gpt4o-mini-smoke --model openai/gpt-4o-mini --thinking off --message "Return exactly: OPENCLAW_AGENT_OK" --json --timeout 180

Observed behavior:

• The GPT-5.4 nano run with --timeout 600 hung past the 10-minute outer shell budget.
• GPT-4.1 mini and GPT-4o mini also hung past their shell timeouts with fresh session IDs.
• One earlier GPT-5.4 nano run with an internal --timeout 180 emitted a command-lane timeout after 210000ms.

Relevant trace from that earlier run:

[agent/embedded] [trace:embedded-run] ... model-resolution:28451ms, auth:37388ms, context-engine:2ms, attempt-dispatch:41880ms
CommandLaneTaskTimeoutError: Command lane "main" task timed out after 210000ms

Expected behavior

For a one-line prompt and a working model/auth profile, agent --local should return a final response or a clear error instead of hanging past the configured timeout and the outer shell budget.

Actual behavior

capability model run --local succeeds for three tested OpenAI models, but agent --local hangs for the same profile and auth. Changing models did not resolve the hang, so the issue appears to be in the native Windows agent/PI command lane rather than the provider/model adapter.

Why this seems useful

This reproduces a Windows hang without any channel plugin in the loop. It separates model inference from the full embedded agent runtime:

• Model/auth path: works
• Full native Windows agent --local: hangs

That should make it easier to debug the command lane or embedded runtime separately from provider credentials and channel-specific issues.