Mention installing openai integration on less powerful devices

[lukasdotcom]

Tested skills on a cloud agent with setting up nextcloud environment, setting up ai, creating php app, and exapp. Mostly got stuck on setting up the ai environment as the computer was not powerful enough for llm2.

Summary

• Add openai-integration.md for installing and configuring integration_openai (API keys, model discovery, modality flags, verify/troubleshoot).
• Point the AI stack skill at that path when hosts lack GPU/disk for stock llm2 multi-GB init (Cloud Agents, CPU-only).
• Harden ai-stack.md for version-skew Assistant installs, in-container --info-xml paths, and required Vite builds on source PHP AI apps.

[oleksandr-nc]

Thanks for this - the openai-integration page fills a real gap and is careful where it matters. I checked the claims against the integration_openai source and a live master instance: all thirteen config keys exist, the php-scoper hook, --sensitive, worker --once, app:remove, the personal-key override and the empty-url default all hold, and the default_speech_model_id catch is a good one. Validators and CI are green.

Inline comments follow. Two are things I think need fixing before merge (the shadowing explanation and file ownership); the rest are suggestions. If you don't see the point in any of them, say so - I'd rather discuss than have you fix things you disagree with.

[oleksandr-nc]

Suggestion, not a blocker: occ app:enable <app> --force skips the version check without editing a tracked file, which also removes the "don't push that bump" caveat. I used it on master today for six store apps and they ran; I have not checked whether occ upgrade re-evaluates it later. Only applies to PHP apps - ExApp manifests still need the local copy.

[lukasdotcom]

--force is probably a nicer way of doing it instead of modifying the app

[oleksandr-nc]

I don't think this is needed: the providers are registered in Application::register() on every request, so flag changes take effect immediately; the only lag is the 60 s task-type cache that ai-stack.md already documents. Disabling an app on a production instance is a heavy way to wait a minute.

[lukasdotcom]

Might be worth it to mention instead clearing the redis cache as that will work without waiting.

[oleksandr-nc]

give approve in advance to unblock merging

[lukasdotcom]

I've retested this again and this time a new issue came up in openai integration of it not reloading the model cache. I've added a script to ensure that this is done and that the model list is correctly seen in the newest commit.

[oleksandr-nc]

Re-tested the updated commands on master with 5.0.0: the store path and the acceptance recipe work, the model-cache catch is a good one. Two commands need a tweak (the admin /models GET answers 412 with basic auth, and composer as www-data fails in the host-owned clone) - merging now and I will send a small follow-up with the verified variants.