[diegosouzapw]

Summary

• Adds NaN.Builders (nan, https://api.nan.builders/v1) as an OpenAI-compatible API-key provider: chat with live /v1/models discovery, embeddings (qwen3-embedding, 4096 dims), STT (whisper), TTS (kokoro), image generation (flux-2-klein, qwen-image-2.1) and rerank (rerank).
• Adds a usage fetcher for GET https://cloud-api.nan.builders/api/usage/quota that maps each per-model {model, tokensUsed, cap, periodEnd} row to a token quota in the limits panel. A malformed report (duplicate model, negative/non-integer counters, cap <= 0, unparseable periodEnd) is reported as unavailable; it is never partially guessed.
• The OpenAI-compatible speech path now reads an optional defaultVoice from the audio provider config, so a Kokoro request without voice sends af_heart instead of the OpenAI stock alloy (which Kokoro does not have). Providers without defaultVoice still fall back to alloy.

Attribution

Thanks to @tasarren for the original implementation.

Changes

• open-sse/config/providers/registry/nan/index.ts (new) + open-sse/config/providers/index.ts: chat registry entry (buildOpenAiCompatibleRegistryEntry, modelsUrl for live discovery, small seed list kept as the offline fallback).
• src/shared/constants/providers/apikey/gateways.ts: display entry (name, color #6D5EF3, website, apiHint pointing to https://cloud.nan.builders). src/shared/constants/config.ts: PROVIDER_ENDPOINTS.nan.
• open-sse/config/embeddingRegistry.ts, audioRegistry.ts (STT + TTS, new optional AudioProvider.defaultVoice), imageRegistryData.ts, rerankRegistry.ts: curated nan entries.
• open-sse/handlers/audioSpeech.ts: voice: body.voice || providerConfig.defaultVoice || "alloy" (one line).
• open-sse/services/usage/nan.ts (new) wired into usage.ts, usage/fetcherProviders.ts, usage/supportedProviders.ts, src/lib/usage/providerLimits.ts and the ProviderLimits label map. Errors go through sanitizeErrorMessage(), there is a 10 s timeout, and the upstream body is never echoed.
• src/app/api/providers/[id]/models/discovery/providerSets.ts: nan added to the named OpenAI-style discovery set.
• docs/reference/PROVIDER_REFERENCE.md: only the total (372 → 373), the API-key section count (249 → 250) and the new nan row.
• Pinned-count tests: provider-node-reserved-prefix (431 → 432, id and alias are both nan) and providers-constants-split (249 → 250).

Not ported (deliberately):

• Built-in /v1/images/edits for nan/flux-2-klein. classifyImageEditTarget in src/app/api/v1/images/edits/route.ts only accepts codex/fal/adobe/syntx/openrouter and custom nodes as built-in edit targets. Supporting NaN edits would change that public route's dispatch, so this PR leaves it out. Generation works. Edits can come in a follow-up.
• The provider icon PNG. The dashboard falls back to textIcon ("NaN").

Provider count docs (expected red)

check:docs-counts (and therefore check:docs-all) is red on purpose. The aggregate "372 providers" prose in README/AGENTS.md/llm.txt/docs/SVGs/package.json description/skills now reads 373. Per the maintainer's decision for this campaign, the count prose and the protected agent-instruction surfaces are not touched here. A single count-sync PR lands at the end of the campaign. The real number also depends on merge order with the sibling provider PR #16889 (DaoXE), which bumps the same counters.

Test plan

• node --import tsx/esm --test tests/unit/nan-provider.test.ts failed before the port (module open-sse/services/usage/nan.ts missing, registries returned null) and passes after: ℹ tests 11 ℹ pass 11 ℹ fail 0
• Related suites (342 test files that reference the provider registries, usage provider lists, modality registries, discovery sets and pinned provider lists), run with the test:unit:ci preamble and an isolated DATA_DIR: ℹ tests 3248 ℹ pass 3244 ℹ fail 4. After updating the two pinned counts, only 2 failures are left:
  • check-docs-counts-sync.test.ts: the expected provider-count drift described above.
  • models-catalog-route.test.ts › "v1 models catalog exposes bare Codex-preferred IDs" (codex-auto-review must not be listed): pre-existing on base. It fails the same way on a pristine origin/release/v3.8.52 checkout (ℹ tests 44 ℹ pass 43 ℹ fail 1) and has nothing to do with nan.
• Re-run of the pinned-count tests: provider-node-reserved-prefix + providers-constants-split pass.
• npx vitest run tests/unit/ui/ProviderIcon-icon-url.test.tsx: Tests 94 passed (94)
• node --import tsx scripts/check/check-provider-consistency.ts: OK — 290 entradas REGISTRY, 373 providers canônicos, 0 exceção(ões) registry-only
• npm run typecheck:core: exit 0
• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files>: clean. npx prettier --check <changed files>: all formatted
• node scripts/check/check-file-size.mjs: OK (no ceiling raised)
• GITHUB_BASE_REF=release/v3.8.52 npm run check:test-masking: OK — 2 modificado(s) … sem enfraquecimento
• check-pr-test-policy.mjs (PR context): Result: PASS
• check:env-doc-sync, check:doc-links, check:fabricated-docs: pass. check:docs-counts: expected red (see above)

live_check (hold-vps)

The endpoints of this vendor have not been verified live. Before merge, on the VPS with a real NaN.Builders key:

# chat + discovery
curl -s $OMNI/v1/models -H "Authorization: Bearer $KEY" | jq '.data[].id' | grep '^"nan/'
curl -s $OMNI/v1/chat/completions -H "Authorization: Bearer $KEY" -H 'content-type: application/json' \
  -d '{"model":"nan/deepseek-v4-flash","messages":[{"role":"user","content":"ping"}]}' | jq .choices[0].message
# embeddings / rerank / tts / image
curl -s $OMNI/v1/embeddings -H "Authorization: Bearer $KEY" -H 'content-type: application/json' \
  -d '{"model":"nan/qwen3-embedding","input":"hi"}' | jq '.data[0].embedding | length'   # expect 4096
curl -s $OMNI/v1/rerank -H "Authorization: Bearer $KEY" -H 'content-type: application/json' \
  -d '{"model":"nan/rerank","query":"q","documents":["a","b"]}' | jq .results
curl -s $OMNI/v1/audio/speech -H "Authorization: Bearer $KEY" -H 'content-type: application/json' \
  -d '{"model":"nan/kokoro","input":"hello"}' -o /tmp/nan.mp3 && file /tmp/nan.mp3
curl -s $OMNI/v1/images/generations -H "Authorization: Bearer $KEY" -H 'content-type: application/json' \
  -d '{"model":"nan/flux-2-klein","prompt":"a red cube"}' | jq '.data | length'

Also confirm that the dashboard Limits panel shows per-model token bars for the NaN connection (quota endpoint cloud-api.nan.builders/api/usage/quota).

⚠️ base-red inherited: #15306 — unit/vitest/integration/package-artifact gates on release/v3.8.52

[github-actions]

CI Coverage Report

• Coverage job: skipped
• PR test policy: success

Coverage artifact was not available for this run.