Version

OpenClaw 2026.5.5 (commit b1abf9d). Reproduced on 2026.5.4 and 2026.4.29 too.
macOS Darwin 25.4.0, node v22 (homebrew).

Summary

The browser plugin (chrome-mcp transport) fails consistently during long agent cron runs with a recurring set of errors that make multi-step browsing flows impossible:

• [tools] browser failed: action targetId must match request targetId
• [tools] browser failed: AbortError: This operation was aborted
• [tools] browser failed: fill requires fields (called with ref + text in raw_params)
• [tools] browser failed: type requires ref or selector
• [tools] browser failed: timed out. Restart the OpenClaw gateway
• [tools] browser failed: Error: ENOENT: no such file or directory, open '/tmp/openclaw/openclaw-chrome-mcp-XXXXXX/<uuid>' (screenshot)

Errors appear across all transports we tried: managed openclaw profile (port 18800), existing-session user profile, and a custom profile with driver: existing-session + dedicated cdpUrl and userDataDir.

Reproduction

1. Cron job, 5-min frequency, --session isolated --light-context --timeout-seconds 300.
2. Agent prompt: open WhatsApp Web tab → snapshot list of unread chats → for each, click → snapshot conversation → type reply → click send.
3. After 1-3 successful tool calls, refs invalidate. Subsequent act calls (click, type, fill) fail with targetId must match request targetId or AbortError.
4. Even a fresh snapshot after the error returns refs that don't survive to the next act.
5. Run hits the 300s timeout. consecutiveErrors climbs to 7+ across runs.

Workarounds tried (do NOT resolve)

• Upgrade 2026.4.29 → 2026.5.4 → 2026.5.5.
• Restart gateway + Chrome managed (launchctl kickstart -k gui/<uid>/ai.openclaw.gateway).
• Reduce open tabs to 3 (WhatsApp Web, IG Direct, a Google Sheet).
• Custom profile with driver: existing-session + cdpUrl + dedicated userDataDir.
• --light-context + --timeout-seconds 300 + housekeeping cron resetting session files when message count > 50 and idle > 10 min.

Impact

Business-critical: the cron supervises sales WhatsApp + Instagram for a store. Failure rate is roughly ~70% per run. Customer messages go unanswered for hours.

Logs (representative excerpt)

[tools] browser failed: action targetId must match request targetId raw_params={"action":"act","target":"host","profile":"openclaw","targetId":"t4","kind":"click","ref":"e229"}
[tools] browser failed: AbortError: This operation was aborted raw_params={"action":"snapshot","target":"host","profile":"openclaw","targetId":"t3","refs":"aria","depth":4,"maxChars":12000}
[tools] browser failed: fill requires fields raw_params={"action":"act","target":"host","profile":"openclaw","targetId":"CD9...","kind":"fill","ref":"e2441","text":"..."}
[tools] browser failed: Error: ENOENT: no such file or directory, open '/tmp/openclaw/openclaw-chrome-mcp-yusK5P/12e7c416-...' raw_params={"action":"screenshot","profile":"aura-wa","target":"host","targetId":"3"}
[diagnostic] stuck session: sessionKey=agent:main:cron:cab08f1d-...:run:... state=processing age=240s reason=processing_without_queue
[agent/embedded] embedded run failover decision: stage=assistant decision=surface_error reason=timeout from=openai-codex/gpt-5.5

Environment

• 3 Chromes running concurrently: managed openclaw on 18800, an ad-hoc Chrome on 19999, and a persistent profile on 20001 (separate userDataDir).
• Plugin @openclaw/[email protected] installed via ClawHub.
• Provider: openai-codex (ChatGPT Pro OAuth).
• 2 agents (main and aura) with separate workspaces and routing bindings.

Question

Is there a newer branch where the chrome-mcp transport's act/click/type/fill validator was rewritten to not lose refs between snapshot and act calls? Is auto-recovery from a targetId mismatch expected at the plugin layer, or is it the agent's responsibility to reissue snapshot and re-derive refs after every error?

Thanks.