[HouMinXi]

Summary

Claude's /v1/models GET now sends limit=1000, so a single request covers the current catalog instead of Anthropic's default page size. xAI API-key connections accept unknown model ids the same way xai-oauth already does. Live OAuth catalog fetch for xai-oauth is on by default (XAI_OAUTH_LIVE_MODEL_DISCOVERY); set the flag to false to keep the frozen seed. New Grok Build OAuth imports persist autoSync: true so the next published grok id shows up without a registry seed bump.

Claude does not get passthroughModels. A 429 on Claude still cools the whole account. There is no after_id loop.

Test plan

• t tests/unit/flagship-0day-discovery.test.ts tests/unit/xai-oauth-discovery.test.ts tests/unit/l1-oauth-autosync-default.test.ts tests/unit/combo-claude-per-model-scope.test.ts tests/unit/feature-flags-doc-sync-static.test.ts tests/unit/grok-cli-oauth.test.ts — 87/87.
• Inject: drop ?limit=1000 from the Claude URL → the exact-URL test fails. Restore → pass.
• Inject: drop passthroughModels on xaiProvider only → xAI apikey half fails, oauth half still green.
• Inject: drop autoSync on mapImportedToken → JWT/auth.json tests fail, browser path still green. Drop it on mapGrokBuildBrowserTokens → the reverse.
• combo-claude-per-model-scope still treats Claude 429 as account-wide.

Notes

xai / xai-oauth / xao stay out of CATALOG_SIBLING_IDS and PROVIDER_SEARCH_PAIRS. Seeds do not add grok-4.7, gemini-4.0-pro, or claude-opus-5-1.

⚠️ base-red inherited: #13866

[HouMinXi]

Adding context on the semantic updates here and the boundaries we kept:

xAI live discovery on by default
XAI_OAUTH_LIVE_MODEL_DISCOVERY was previously off by default, leaving xai-oauth users on static seed lists unless they knew to export the flag. Since xAI rotates model names regularly, seeds go stale quickly. The live discovery path in route.ts already falls back to the static catalog if xAI is unreachable, so flipping the default gives users fresh chat IDs without breaking offline or degraded startups.

autoSync on new Grok Build imports
Browser and auth.json token imports used to omit autoSync, so imported accounts never updated their model list unless re-imported from scratch. Stamping autoSync: true on import lets the existing sync loop keep them current.

xAI API key model passthrough
We enabled passthroughModels for xAI API keys to match xai-oauth. Unrecognized model IDs pass through to xAI rather than getting blocked by our local schema validator. If an ID is invalid, xAI returns 404, and perModelFailureScope already isolates that failure to the single model ID.

Why Claude was intentionally excluded from passthrough
We specifically did not enable passthroughModels on Claude. Enabling passthrough on Claude turns hasPerModelQuota("claude") on, which changes 429 handling from cooling the entire account to locking just the failed model name. That would let combos keep pounding an already-exhausted Anthropic key with sister models. For Claude, requesting limit=1000 on /v1/models gets the full current catalog in a single call without touching the quota failure domain.

[HouMinXi]

Merge-order note for whoever lands second between this PR and #14525.

#14525 (grok build-object supportedThinkingEfforts) intentionally seeds grok-4.7 into the registry and keeps xhigh on grok models discovered from object catalogs. If it merges first, two guard tests in this PR need updating:

• tests/unit/flagship-0day-discovery.test.ts, test 9 ("zero-day models absent from support surface"): grok-4.7 has to come out of the banned list, since it will be present on purpose.
• tests/unit/provider-models-discovery-split.test.ts, "preserves exact supported reasoning efforts": grok now keeps xhigh, so the expected lists for 4.5/4.6/4.7 gain "xhigh", and the third fixture model (grok-4.8, object form with only xhigh declared) yields ["xhigh"] instead of undefined under the explicit-list filtering semantics.

If this one merges first instead, #14525 needs the reciprocal updates on its side. I hit both failures while building an integration branch carrying the two PRs together, so this is measured, not speculative.