ik this is still a draft but why do you want to evaluate getHfEndpoint everytime? Why can't we just do it once and save the result

Pushed commit 40e59750 addressing review feedback. Summary of fixes:

hf_endpoint.py — datasets-server fallback: get_hf_datasets_server() no longer returns the Hub mirror URL when HF_DATASETS_SERVER is unset. Most Hub mirrors don't proxy /splits, so the previous fallback turned every dataset-splits lookup into a silent 404. New behavior: return HF_DATASETS_SERVER if set, otherwise the official datasets-server.huggingface.co. Added a startup WARNING when HF_ENDPOINT is mirrored but HF_DATASETS_SERVER is unset, so operators know they need to opt in explicitly.

instruction-from-answer.json — hardcoded endpoint: Removed "endpoint": "https://huggingface.co". parseSeedSettings reads sourceRaw.endpoint and that value was overriding the getHfEndpoint() default in the parser, so a user on a mirror who loaded this template ended up with the wrong endpoint pre-filled in their seed config.

hf-endpoint.ts + env.ts — cache → direct store reads: Removed the module-level cache and hydrateHfEndpoints(); both getters now read usePlatformStore.getState() on every call. This:

1. Fixes a pre-hydration race where code (e.g. makeSeedConfig, parseSeedSettings) that ran before /api/health resolved would bake the default endpoint into new configs.
2. Drops the circular import between env.ts and hf-endpoint.ts.
3. getState() is O(1) — no measurable perf cost compared to the cache.

env.ts — /api/health catch branch: Widened the warning message to state explicitly that HF_ENDPOINT configuration will not be applied when the health fetch fails, so mirror-routing issues surface in the console rather than looking like a generic network blip.

transformers_version.py — log severity: _check_tokenizer_config_needs_v5 and _check_config_needs_550 now distinguish HTTPError 404 (DEBUG, legitimate fail-open) from HTTPError 5xx and URLError (WARNING with a hint about HF_ENDPOINT). Previously a mirror returning 500/connection-refused looked identical to a legitimate 404, so mirror problems surfaced as cryptic transformers crashes with no log trail.

@Datta0 your concern about evaluating getHfEndpoint on every call was valid — commit 40e59750 actually undid the caching (aaddd281) that addressed it, which was a mistake on my part. Restored in ae620853:

let _endpoint = usePlatformStore.getState().hfEndpoint || DEFAULT_HF_ENDPOINT;

usePlatformStore.subscribe((state) => {
  _endpoint = state.hfEndpoint || DEFAULT_HF_ENDPOINT;
});

export function getHfEndpoint(): string {
  return _endpoint;
}

Getters now read a module-level variable (single property lookup, no getState() cost), and a Zustand subscribe keeps the cache in sync whenever /api/health populates the platform store. Compared to the original hydrateHfEndpoints() approach this also removes the circular import with env.ts and the pre-hydration window where a stale default could leak into factories.

Storage-wise everything stays in-memory: usePlatformStore is a plain (non-persist) store and the module variables live in module scope, so each page reload re-fetches /api/health and the backend's live HF_ENDPOINT env var wins.

pytest tests successfully passed

other tests
Default — no env vars (regression baseline)

unset HF_ENDPOINT HF_DATASETS_SERVER
  # start studio as usual, then:
  curl -s http://localhost:<port>/api/health | jq '{hf_endpoint, hf_datasets_server}'
  # expect: { "hf_endpoint": "https://huggingface.co",
  #           "hf_datasets_server": "https://datasets-server.huggingface.co" }
  Then exercise model search, dataset search, dataset splits in the UI — all should work as before.

Other Tests mirror via hf-mirror.com

export HF_ENDPOINT=https://hf-mirror.com                                                        
# start studio
curl -s http://localhost:<port>/api/health | jq '{hf_endpoint, hf_datasets_server}'             
# expect: { "hf_endpoint": "https://hf-mirror.com",                                             #           "hf_datasets_server": "https://datasets-server.huggingface.co" }
Backend logs should contain the new startup WARNING from hf_endpoint.py:
  HF_ENDPOINT is set to https://hf-mirror.com but HF_DATASETS_SERVER is unset; datasets-server c
alls will still go to https://datasets-server.huggingface.co. Set HF_DATASETS_SERVER to
  override.

  In the browser devtools Network tab:
  - Model search → requests hit hf-mirror.com/api/models ✓
  - Dataset search → requests hit hf-mirror.com/api/datasets ✓


  A new seed node created in recipe-studio should pre-fill hf_endpoint: https://hf-mirror.com (n
ot huggingface.co).