Quick context: I'm using OpenClaw with
Gemini CLI as the backend. I found that the
harness (bootstrap with IDENTITY/USER/SOUL
files) never reaches the headless Gemini
CLI (gemini --prompt) because there's no
systemInstruction channel β it's
spawn-based and OpenClaw only controls
--prompt.
I migrated to an ACP binding with gemini --acp on a Telegram supergroup (topic:9).
The hypothesis is that ACP mode, being a
persistent bidirectional protocol, has a
separate channel through which OpenClaw can
deliver the bootstrap context β not just
the user text via --prompt.
Two questions:
-
Does
gemini --acpactually receive the
bootstrap (IDENTITY/USER/SOUL) via ACP?
Or does the harness gap exist in ACP mode as
well? -
After configuring the binding and
applying an ACP session reset (kill orphan
processes + clean state files + gateway
restart),/newin the topic responds
normally β but a regular message produces no
reply and no "Typing..." indicator in
Telegram. What's the right diagnostic path
to determine whether this is model quota
silencing the dispatch or something else
happening before it even reaches the
channel?