utils.llm.call_llm silently falls back to fake provider — Copilot provider never registered at boot

[kody-w]

Summary

utils/llm.py exposes a clean call_llm() helper that's meant to let plug-in agents share the brainstem's authenticated Copilot session. But brainstem.py never calls register_copilot_provider() at boot, so _copilot_token_provider stays None, detect_provider() falls through to chat_fake, and any agent that calls call_llm gets the prompt echoed back as text instead of an LLM response.

How to reproduce

1. Start the brainstem normally (python brainstem.py).
2. From any plugin agent, call:

   from utils.llm import call_llm
   print(call_llm([{"role":"user","content":"say hi"}]))

3. Observe output: fake-llm: say hi instead of a real model response.

Root cause

utils/llm.py line ~44–55 defines register_copilot_provider(token_getter, default_model) and a module-level _copilot_token_provider = None. detect_provider() only returns "copilot" if that variable is set. Nothing in brainstem.py ever sets it — brainstem.py calls its own internal call_copilot() directly.

Impact

This is the silent killer behind the second issue I'm about to file (LearnNew shelling out to a nonexistent CLI). Every plug-in agent that wanted to share the brainstem's session is hitting the fake provider without warning. The fake response (fake-llm: <prompt>) is then injected into generated code as if it were real, producing files with the prompt baked into the function body as a syntax error.

Suggested fix

In brainstem.py at the boot section (after load_soul() and before load_agents()), register the provider:

try:
    from utils.llm import register_copilot_provider as _reg_copilot

    def _llm_token_getter():
        tok, ep = get_copilot_token()
        return tok, ep

    _reg_copilot(_llm_token_getter, default_model=MODEL)
    _tlog("llm.copilot_registered", {"model": MODEL})
except Exception as _llm_err:
    _tlog("llm.copilot_register_failed", {"error": str(_llm_err)}, level="warn")

Also call it inside set_model() so the registered default tracks runtime model switches.

Environment

• RAPP brainstem 0.15.9
• macOS 14, Python 3.11 venv
• Copilot session authenticated via device code

I patched this locally and confirmed LearnNew's body generator (issue #2) starts producing real code afterward.