feat(llm): add native Novita AI provider via OpenAI-compatible endpoint

[jax-novita]

Contribution Receipt

• Change: Add Novita AI (https://api.novita.ai/openai/v1) as a first-class LLM provider, following the same pattern used for DeepSeek/HuggingFace/NanoGPT — dispatched through the existing OpenAIAdapter, resolved via NOVITA_API_KEY, with a default model, curated model catalog entries, config/setup wiring, and regression tests.
• Scope class: local_lab
• Target authority: n/a — LLM provider integration, no target scope involved
• Network use: authorized_external (live call to api.novita.ai/openai/v1 to verify catalog and chat responses during development; no network calls in the shipped test suite)
• Run mode labels: n/a
• Model/harness labels: n/a
• Commands run:
  • npm run typecheck -> pass
  • npx vitest run src -> 757/758 pass (the sole failure is a pre-existing environment-dependent local-agent-path-resolution.test.ts assertion tied to codex being on the local PATH; unrelated to this change)
  • npm run doctor -> pass, 0 blockers
• Artifacts: none
• Redaction: none required — no secrets or credentials touch the diff; NOVITA_API_KEY follows the existing .env.example pattern used by other providers
• Claims changed: none
• Abstentions/refusals: none
• Residual risk: none — change is additive and follows the existing provider-registration pattern; no existing provider behavior is modified

Verification

• git diff --name-status upstream/main...HEAD reviewed: only Novita provider wiring and its tests are touched (.env.example, src/config/index.ts, src/config/provider-models.ts, src/llm/index.ts, src/setup.ts, src/types/index.ts, and new/updated tests). No unrelated deletions.
• Live smoke test through createNovitaBackbone() → LLMBackbone.chat() → OpenAIAdapter against the authenticated Novita catalog confirmed a working chat completion (deepseek/deepseek-v4-pro-0813, finishReason: "stop", usage populated).

[jmagly]

Reviewed exact head 6a1af1b31967b6d6f09ce9a8f90fc2b6a28f4c84 against main at 6d4e017, including the hosted-provider trust boundary.

The service and wire shape are real: Novita's current official documentation corroborates the OpenAI-compatible base, Bearer authentication, /openai/v1/models, and /openai/v1/chat/completions; its current public catalog lists GLM 5.2 and DeepSeek V4 Pro 0813 with the dimensions used here. The focused config/model-list/redaction tests pass.

Two blockers remain:

1. The named-provider integration is incomplete across user-facing contracts. README.md, docs/GETTING_STARTED.md, and the provider controls/model lists in docs/index.html still omit Novita, so a key can be configured through the terminal wizard but the documented and browser setup/status surfaces disagree. Add the provider to those surfaces and disclose that prompts, outputs, request metadata, and the API credential leave T3MP3ST for a third-party hosted service. Link the current provider privacy/terms material without making an unsupported sensitive-workload assurance claim.

2. The tests prove registry/config wiring and model discovery only. Add provider-specific mocked HTTP regressions through the actual backbone for the correct chat URL and authorization, success, 401/429, timeout/malformed response, and no unintended fallback. The PR currently reports “residual risk: none,” but remote-provider retention, subprocessors, availability, and model-catalog drift remain external unknowns; record those honestly.

Verification:

• npm ci: pass
• npm run typecheck: pass
• focused Novita/provider-model/redaction tests: 28/28 pass
• full local suite: 757/758; only the known environment-dependent Codex PATH assertion failed, as already disclosed in the PR
• hosted checks: none reported for this head

Provider verdict: service real; wire/catalog partially independently verified; sensitive-workload trust unproven; integration incomplete; not merge-ready.

[jmagly]

I completed the remaining bounded review work as a maintainer courtesy and pushed signed commit c80ec42 to this branch after merging current main.

What I changed:

• added Novita to the browser Universal API Config and its canonical base URL;
• documented the hosted data flow and linked Novita's current privacy policy/terms without asserting sensitive-workload assurance;
• added a provider-specific adapter so missing credentials name NOVITA_API_KEY rather than an OpenAI key;
• propagated OpenAI-compatible HTTP status as LLMApiError, including bounded Retry-After handling;
• added real-backbone mocked wire tests for URL, Bearer auth, request model/body, 401/429 behavior, and opt-in-only cross-provider fallback;
• synced the published docs source.

Verification on the prospective merged tree:

• npm run typecheck: pass
• focused provider/model/redaction/UI tests: 36/36 pass
• npm test: 767/767 plus ops/model/refusal gates pass

Thank you for the substantive provider implementation. I’ll recheck hosted CI and the exact new head before merging.

[jmagly]

Re-reviewed exact head c80ec425b2d35846309094a4798cb03f5d746ff7 against current main at 41f0c140397ac41e09d03cb255d77540e61a9c00.

The previously requested integration-completeness and provider-boundary changes are resolved. Official Novita documentation corroborates the OpenAI-compatible endpoint, Bearer authentication, model-list route, and current catalog entries. The user-facing setup surfaces now agree, hosted data flow and residual trust limits are disclosed, missing credentials identify NOVITA_API_KEY, HTTP status is preserved for retry/fallback policy, and focused tests cover the actual chat wire plus 401/429 and opt-in-only fallback.

Verification:

• signed maintainer courtesy commit verified
• npm run typecheck: pass
• focused provider/model/redaction/UI tests: 36/36 pass
• npm test: 767/767 plus ops/model/refusal gates pass
• hosted exact-head CI: pass

Provider verdict: service real; endpoint/catalog independently corroborated; sensitive-workload trust remains explicitly unproven and disclosed; integration complete for the claimed scope; merge-ready.

Decision: approve this exact head. Review expires on head/base/check drift.