fix(ai): replace decommissioned Groq model IDs

[Mr-Macharia]

Summary

Groq is currently unusable — all four model IDs in PROVIDER_MODELS.groq have been decommissioned by Groq, so every request fails with model_not_found.

Verified against the live Groq API. The previous default:

POST https://api.groq.com/openai/v1/chat/completions
{"model": "meta-llama/llama-4-scout-17b-16e-instruct", ...}

→ {"error": {"message": "The model `meta-llama/llama-4-scout-17b-16e-instruct`
   does not exist or you do not have access to it.",
   "type": "invalid_request_error", "code": "model_not_found"}}

GET /openai/v1/models returns 14 models, and none of the four configured IDs are among them.

Changes

src/config/aiProviders.ts — replaced the four retired models. Each replacement was smoke-tested against /v1/chat/completions and returns HTTP 200:

Model	Category
openai/gpt-oss-120b	Reasoning · new default
openai/gpt-oss-20b	Speed
qwen/qwen3.8-27b	Reasoning
groq/compound	Performance

Of the 14 models Groq returns, only 6 are chat models — the rest are speech-to-text (whisper-*), text-to-speech (canopylabs/orpheus-*), and safety classifiers (llama-prompt-guard-2-*, gpt-oss-safeguard-20b). These four are the general-purpose chat models best suited to diagram generation.

14 locale files (src/i18n/locales/ and public/locales/, 7 languages each) — model IDs double as i18n keys. AISettings.tsx:226-229 looks up settingsModal.ai.models.<provider>.<translateKey>.{label,hint,category,badge}. Changing a model ID without the matching i18n entry makes the dropdown render raw key paths instead of names, so these move together.

README.md — the provider table listed the retired model as Groq's default.

Test plan

• npx tsc --noEmit — clean
• npm run lint — clean
• npm test — 295 files, 1445 tests passing
• Each new model ID returns HTTP 200 from Groq's /v1/chat/completions
• Old default confirmed returning model_not_found
• src/i18n/locales/ and public/locales/ verified byte-identical after the change
• JSON round-trip verified lossless, so the diff touches only the groq block
• Manual: Settings → AI → Groq, confirm the dropdown shows model names (not raw keys) and a generation succeeds

Notes for reviewers

Machine-translated strings. The new label/hint/category/badge strings for de, es, fr, ja, tr and zh were written without native-speaker review. They follow the tone of the surrounding entries, but a speaker should sanity-check them. Happy to revert any locale to English placeholders if you'd rather handle translation separately.

NVIDIA is probably affected too, and is not fixed here. PROVIDER_MODELS.nvidia still lists meta/llama-4-scout-17b-16e-instruct and meta/llama-4-maverick-17b-128e-instruct — the same retired Llama 4 generation under NVIDIA's naming. Different provider and lifecycle, so I did not assume it's broken and had no NVIDIA key to test with. Worth a follow-up issue.

Underlying cause. Hardcoded model lists go stale silently — the user just sees a failed request with no hint that the model no longer exists. Fetching /v1/models at runtime with the static list as offline fallback would make this self-healing across every OpenAI-compatible provider. It needs a curation layer (Groq's endpoint returns Whisper and classifier models that don't belong in a diagram-model dropdown), so it's a feature rather than a fix and is deliberately out of scope. Glad to open an issue if there's interest.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.