Add Nous Research provider

[Acorzo1983]

First pass: adds the Nous Research provider, compliant logo, and four requested stable models: Hermes 4 70B, Hermes 4 405B, Claude Haiku 4.5, and Kimi K3. The catalog sync and routed-model expansion remain deferred for a later, separately reviewed change. Source for availability, provider pricing, limits, modalities, and supported parameters: https://inference-api.nousresearch.com/v1/models. Nous does not document provider-specific request syntax for reasoning controls, so all reasoning models use empty reasoning_options. Validation: bun validate and git diff --check pass.

[github-actions]

Action items

• [critical] [violation]providers/nous/provider.toml:1 - Check: New providers must ship a compliant providers/<id>/logo.svg. Why: This PR adds provider nous with no logo file. AGENTS.md treats a missing compliant logo as a merge blocker. Action: Add providers/nous/logo.svg using currentColor, no fixed size or hardcoded colors, preferably a square viewBox.
• [high] [possible mistake]providers/nous/models/nousresearch/hermes-4-405b.toml:1 - Check:reasoning_options must describe controls exposed by this inference provider, not another gateway. Why: Both models claim type = "toggle", but the PR says capabilities were mirrored from OpenRouter and does not identify a Nous request field/values that enable and disable reasoning on the same model ID. Toggle requires a verified caller control; otherwise use reasoning_options = []. Action: Verify the Nous API toggle (cite path/values or reproduce), keep toggle only with that evidence (and the exact request syntax comment if documented), or set reasoning_options = [] on both models.
• [low] [possible mistake]providers/nous/models/nousresearch/hermes-4-70b.toml:16 - Check: Material model data should be justified with first-party provider sources. Why: The PR states costs, limits, and capabilities mirror OpenRouter rather than Nous Research pricing/docs. That can silently publish another provider’s prices and feature flags for nous. Action: Cite Nous-specific pricing and capability docs (or API evidence) for cost, limits, tool_call, and related fields, and correct any values that differ from OpenRouter.

[Acorzo1983]

Closing until the Nous OpenCode integration has an authenticated local smoke test. I will reopen or replace this PR after verifying inference and agent behavior.

[github-actions]

Action items

• [high] [possible mistake]packages/core/src/sync/providers/nous.ts:36 - Check:reasoning_options must describe controls on Nous’s inference API, not another gateway’s catalog. Why:translateModel reuses buildOpenRouterModel, which maps OpenRouter-shaped reasoning.supported_efforts / supports_max_tokens into dozens of non-empty effort and budget_tokens entries (e.g. openai/gpt-5-mini, anthropic/claude-opus-4.7-fast, meituan/longcat-2.0). The PR only cites a generic API-docs URL and OpenRouter mirroring; it does not show that Nous’s chat/completions surface accepts those controls (OpenRouter even documents route-specific mappings such as Opus 4.7 Fast via verbosity). Unverified options mislead clients. Action: Until each control is verified on Nous, emit reasoning_options = [] for reasoning models (as Chutes does when the catalog only advertises capability). Only keep non-empty options with Nous request-field evidence in the PR body; adjust the sync (and tests) so OpenRouter reasoning metadata is not treated as Nous controls by default.
• [high] [possible mistake]providers/nous/models/nousresearch/hermes-4-70b.toml:15 - Check:toggle requires a verified same-ID on/off control on this provider. Why: Both Hermes models claim type = "toggle" with no adjacent API syntax comment and no Nous citation for the request path/values. The first commit matched OpenRouter’s Hermes files; the sync test then preserves that authored toggle. That does not meet the audit standard (“complete the toggle sentence with provider request path”). Action: Set reasoning_options = [] for both Hermes models unless you document and cite the exact Nous request field/values that enable and disable reasoning, and stop the sync/test from re-preserving an unverified toggle.
• [low] [possible mistake]providers/nous/models/nousresearch/hermes-4-70b.toml:18 - Check: Data-changing PRs should cite first-party sources mapped to claims. Why: The PR body says Hermes limits/prices “mirror” OpenRouter, but the synced files use different costs (0.05/0.2 and 0.09/0.37 vs OpenRouter 0.13/0.4 and 1/3), and the bulk catalog has no Nous pricing page or claim→URL mapping. Action: Update the PR body with direct Nous pricing/docs links and state what each supports (Hermes prices, other model costs, and that /v1/models is authoritative for the synced set).

[rekram1-node]

i see like no reasoning options anywhere, this is likely wrong.

[rekram1-node]

Also this PR is massive, prefer u do a first pass instead of 2.5k lines in 1

[Acorzo1983]

Reduced this to a first-pass provider PR: removed the catalog sync and routed-model expansion, added the required logo, and retained only Hermes 4 70B/405B. I replaced the unverified reasoning toggles with reasoning_options = [] and added first-party /v1/models citations plus API-derived pricing/context/modalities. Thanks for the guidance.

[github-actions]

No actionable findings.

[Acorzo1983]

Following the first-pass guidance, this remains a small hand-authored PR. I added two requested stable models, Claude Haiku 4.5 and Kimi K3, bringing the total to four models. Their Nous-specific costs/limits are sourced from /v1/models; reasoning_options remain empty because Nous does not document request syntax for the advertised controls.

[github-actions]

No actionable findings.

[rekram1-node]

So you are telling me this provider doesnt support reasoning options at all? All these responses read like an llm so im not sure if you are telling me this or ur agent is so I want to confirm.

[amrrs]

@rekram1-node on a totally different note, any chance you could ask someone from open code to look into this - anomalyco/opencode#38709

[github-actions]

Closing this pull request as stale because it has not been updated in 30 days. Feel free to reopen it or submit a new pull request if the work is resumed.