Summary

ACP workflow in OpenClaw could start runs, but I could not reliably observe/completion-check them from the parent session due visibility restrictions. In the same flow, Chrome relay profile expectations also mismatched available browser profiles.

Impact

• ACP task execution can be launched, but parent session cannot verify completion/details.
• User-facing status updates become ambiguous ("run launched" but no readable result stream).
• Browser relay instructions expecting profile=chrome failed in this environment.

Environment

• Runtime: OpenClaw (Telegram group session)
• Model: codex (openai-codex/gpt-5.3-codex)
• Browser profiles reported: niemand, mac, openclaw

Reproduction

1. Spawn ACP run:

{"runtime":"acp","agentId":"codex","mode":"run","task":"..."}

2. Attempt to read child session output/history from parent session.
3. Attempt browser relay tab attach with profile="chrome" per policy guidance.

Actual behavior

A) ACP session output visibility blocked

session_spawn accepted the run:

{
  "status": "accepted",
  "childSessionKey": "agent:codex:acp:8dd69e20-35ca-4daf-8a19-5c4b696b1f78",
  "runId": "012bbedf-018c-436c-8e2e-afdd2d7c6e44",
  "mode": "run",
  "note": "initial ACP task queued in isolated session; follow-ups continue in the bound thread."
}

But reading history failed:

{
  "status": "forbidden",
  "error": "Session history visibility is restricted. Set tools.sessions.visibility=all to allow cross-agent access."
}

B) Browser relay profile mismatch

Attempt with profile="chrome" failed:

{
  "status": "error",
  "error": "Error: Profile \"chrome\" not found. Available profiles: niemand, mac, openclaw"
}

Expected behavior

• Parent session should have a supported/consistent way to observe ACP run completion + result payload (or an explicit completion callback channel) without requiring global visibility escalation.
• Browser relay guidance should map to available profile names or provide aliasing/clear fallback when chrome is unavailable.

Suggested fixes

1. ACP observability
   • Add a first-class sessions_wait / sessions_result for spawned ACP runs that does not require global sessions.visibility=all.
   • Or emit completion payload to parent session automatically for mode=run.
2. Profile aliasing
   • Support chrome alias if relay is connected to a non-default profile, or expose "relay-attached profile" in browser status.
3. Error UX
   • On forbidden history read, provide actionable alternative call path for parent-safe completion checks.

Notes

As a workaround, I switched to direct coding-agent workflow (Codex PTY) for implementation work, but that bypasses intended ACP orchestration visibility.