#5-9.beta-1 Acpx feedback

1 messages · Page 1 of 1 (latest)

indigo pewter
#

I've been trying to get it working and now using Codex to dig into it because it's gotten closer but still issues

Yes: I checked upstream openclaw/openclaw main just now, commit 0f8b52f9, and I do not see this fixed there.

The ACP path still does this:

OpenClaw runtime options map thinking to ACP config key thinking: runtime-options.ts (line 316)

The ACP control plane then rejects keys not advertised by the backend: manager.runtime-controls.ts (line 64)

Claude ACP advertises effort, not thinking, so OpenClaw’s thinkingDefault does not translate correctly.

There is a Claude thinking-related fix in upstream, but it’s for the Claude CLI backend passing --effort, not for Claude ACP via acpx. That’s the subtle trap here.

So yes, best fix belongs in openclaw/openclaw: its ACP control plane should map OpenClaw’s abstract thinking option to the backend’s advertised key, probably effort for Claude ACP. An acpx-side compatibility alias could also paper over it, but OpenClaw owns thinkingDefault, so OpenClaw is the right architectural place.

wicked rapidsBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> âś… Mark Solution

waxen wagonBOT
#

Yeah, that diagnosis looks right. I just checked current upstream main too — it’s now at df8ccb00, and the same mismatch is still there:

  • src/acp/control-plane/runtime-options.ts still emits OpenClaw thinking as ACP config key thinking
  • src/acp/control-plane/manager.runtime-controls.ts rejects keys that the backend didn’t advertise
  • the inverse parser recognizes reasoning_effort, but not plain effort, and the outbound path doesn’t translate Claude ACP’s effort

So this is not the Claude CLI --effort fix; it’s specifically the ACP/acpx runtime-control mapping path.

Best short-term workaround: don’t set thinkingDefault / /acp set thinking ... for the Claude ACP session. If you need to experiment, try clearing the session runtime options with /acp reset-options, then use the backend-native key directly: /acp set effort high — but if a thinking option is still present, OpenClaw will still try to apply thinking first and fail.

Best upstream fix: OpenClaw should translate its abstract thinking runtime option to whatever ACP config key the backend advertises — probably prefer thinking if supported, else effort, maybe also reasoning_effort/thought_level for adapters that expose those.

If you file this, include your openclaw --version, the upstream commit you checked, and the error showing config key "thinking" rejected. GitHub issue link: https://github.com/openclaw/openclaw/issues

ACP docs for context: https://docs.openclaw.ai/tools/acp-agents

indigo pewter