LLM cannot be initialized from an editor script: DontDestroyOnLoad in Awake()

[merrymaker14]

Describe the bug

Describe the bug

LLM.Awake() calls DontDestroyOnLoad, which Unity forbids outside Play mode. This makes it
impossible to initialize an LLM from an editor script.

The model itself loads and the server starts — the console shows model loaded — and then
Awake() throws:

InvalidOperationException: DontDestroyOnLoad can only be used in play mode and, as such,
cannot be part of an editor script.
LLMUnity.LLM.Awake () at Assets/LLMUnity/Runtime/LLM.cs:365

Because the exception interrupts Awake() before the native caller is created, every request
afterwards fails with LLM caller not initialized.

Why this matters

I'm building editor tooling that lets a developer verify their setup — model present, correct,
and actually producing a reply — without entering Play mode. The rest of my stack (speech
synthesis, model management, validation) works fine from the editor; this is the only thing
that blocks the dialogue half.

Suggested fix

Guarding the call would be enough:

if (Application.isPlaying) DontDestroyOnLoad(transform.root.gameObject);

Alternatively, expose it as the serialized setting that #152 already suggested — "This could
also be a setting in the component itself" — which covers both cases and changes the default
behaviour for nobody.

To be clear: I understand DontDestroyOnLoad was added deliberately (#152, #174) and I'm not
asking for it to be removed. Only for it to be skippable when it cannot legally run.

Steps to reproduce

No response

LLMUnity version

3.0.3

Operating System

Windows

[merrymaker14]

Correction to my own report — the first half of this issue was based on a wrong premise.

The setting already exists. LLM.cs:89 declares public bool dontDestroyOnLoad = true; and
LLM.cs:365 guards the call with it. Setting it to false does remove the exception, exactly as I was
asking for. I missed it because Awake is async: the exception surfaced after the await, so my
try/catch never saw it — the console showed the error while my calling code was told the call had
succeeded. Apologies for the noise.

What actually blocks editor-time initialization is elsewhere, and I don't think it's fixable in the
way I suggested. With the flag off, Awake completes cleanly:

llm.started = True
llm.failed  = False
"LLM service created successfully"

but the native caller is still never built:

llmAgent = null    // after both Awake() and Start()

LLMClient creates it in Start(), which is async void awaiting SetupCaller(). The continuation is
posted to Unity's synchronization context, and that context is only pumped inside the game loop — outside
Play mode it never runs. So this isn't a guard that can be added; it's what "not playing" means. I've
stopped pursuing it and my tooling now verifies speech synthesis from the editor and dialogue in Play mode,
which is fine.

One thing that may be worth keeping, independent of the editor case. Because Awake is async void,
a failure inside it cannot be observed by the caller: the exception goes to the console, Awake returns as
if it succeeded, and every subsequent request fails with LLM caller not initialized. That's not specific
to DontDestroyOnLoad — a bad model path, a missing file or an out-of-memory condition would present the
same way, and the reported symptom is several steps removed from the cause. It cost me a fair amount of
debugging time, and it might be doing the same to others.

If that's worth tracking I'm happy to open it as a separate issue with a minimal repro. Otherwise feel free
to close this one — the original request is already implemented and the rest is out of scope.

Thanks for the library, and sorry again for the incorrect first report.