OpenClaw browser action bug repro, 2026-04-21

Summary

On this macOS host, browser inspection calls work, but browser action calls hang and then poison the browser tool until the gateway is restarted.

This reproduces on:

• OpenClaw-managed browser profile (profile: "openclaw")
• Existing-session Chrome attach (driver: "existing-session", profile aliases including user / chrome-live)
• ChatGPT (https://chatgpt.com/)
• Plain https://example.com/

So this does not look site-specific, auth-specific, or ChatGPT-specific.

Environment

• Host: Kimi’s Mac mini
• OS: Darwin 24.6.0 arm64
• Node: v25.8.1
• OpenClaw runtime in session metadata: agent=main | host=Kimi’s Mac mini | repo=/Users/kimi/.openclaw/workspace
• Browser executable detected: /Applications/Google Chrome.app/Contents/MacOS/Google Chrome
• OpenClaw managed browser CDP: http://127.0.0.1:18800

Repro steps

Case A, OpenClaw-managed browser

1. Restart gateway.
2. Run browser status for profile: "openclaw".
3. Open https://example.com/ in profile: "openclaw".
4. Take a snapshot.
5. Attempt a single click action on the Learn more link.

Observed

• status succeeds
• open succeeds
• snapshot succeeds
• act(kind="click") times out
• after that, browser calls often continue failing until gateway restart

Concrete example

Snapshot on https://example.com/ returned a normal tree including:

• RootWebArea: Example Domain
• link ref ax12, name Learn more

Then this action failed:

{
  "action": "act",
  "target": "host",
  "profile": "openclaw",
  "targetId": "802856F24F36CE8DB2B7BD1ED64890F6",
  "ref": "ax12",
  "kind": "click",
  "timeoutMs": 30000
}

Error returned:

timed out. Restart the OpenClaw gateway (OpenClaw.app menubar, or `openclaw gateway`). Do NOT retry the browser tool — it will keep failing. Use an alternative approach or inform the user that the browser is currently unavailable.

Case B, ChatGPT page

After clean restart, the same pattern occurred on https://chatgpt.com/:

• page load ok
• snapshot ok
• single click into the prompt box timed out immediately

So the failure is not specific to ChatGPT image generation.

Why this seems like an OpenClaw browser-action-path bug

I checked the local compiled code paths.

Relevant code paths

• Browser client timeout wrapping:
  • /opt/homebrew/lib/node_modules/openclaw/dist/session-tab-registry-C-F_cimS.js
• Local browser dispatch runtime:
  • /opt/homebrew/lib/node_modules/openclaw/dist/local-dispatch.runtime-DFCUCw4c.js
• Browser route dispatcher and service bootstrap:
  • /opt/homebrew/lib/node_modules/openclaw/dist/control-service-C6DCgn50.js
• /act route handling:
  • /opt/homebrew/lib/node_modules/openclaw/dist/server-context-qgfDaRbE.js

Important observation

The two action backends appear different:

• existing-session path uses Chrome MCP element actions
• managed openclaw path uses Playwright action execution

But both hang in practice.

That suggests the problem may be in a shared layer, likely one of:

• route/tab context setup
• browser runtime state / request dispatch
• request cancellation / signal handling
• navigation guard / tab resolution around action execution
• some host-specific deadlock in the browser control service

Log evidence

Useful log locations on this host:

• /Users/kimi/.openclaw/logs/gateway.log
• /Users/kimi/.openclaw/logs/gateway.err.log
• /tmp/openclaw/openclaw-2026-04-21.log

Observed behavior in logs:

• repeated top-level browser tool timeout messages
• no useful low-level browser exception emitted alongside the hang
• browser service starts cleanly
• reads/snapshots work before first action failure

Example top-level failure text in tmp log:

[tools] browser failed: timed out. Restart the OpenClaw gateway (OpenClaw.app menubar, or `openclaw gateway`). Do NOT retry the browser tool — it will keep failing. Use an alternative approach or inform the user that the browser is currently unavailable.

Current practical workaround

• Use browser snapshots/read-only inspection only
• Avoid browser act calls on this host/build
• Restart gateway if browser tool gets poisoned
• For actual work, use manual browser interaction with the agent guiding/reviewing

Minimal expected result

A single click on example.com should either:

• succeed, or
• fail with an actionable internal browser error

It should not silently hang for 30s and then poison subsequent browser usage.

One-line repro

Fresh gateway restart -> open example.com in profile=openclaw -> snapshot ok -> single act(kind=click, ref=ax12) hangs for 30s -> browser tool becomes unreliable until restart.