feat: add EUrouter provider and automated model sync

[DavidLoDico]

Summary

EUrouter is the european AI gateway.

• add EUrouter as an OpenAI-compatible provider with the official theme-safe EU logo
• add 122 verified EUrouter model TOMLs from the live catalog
• add an automated sync adapter and register it in the provider matrix / aggregator group
• cover price conversion, modalities, capabilities, reasoning controls, canonical factoring, preservation, and macOS case-sensitive canonical lookup regressions

Integration details

• API base: https://api.eurouter.ai/api/v1
• Catalog: GET /models is public; EUROUTER_API_KEY is optional. On 2026-07-12, public and authenticated requests both returned the same 122-model response, so scheduled sync does not require a new secret.
• Pricing: EUrouter exposes per-token strings with an explicit currency. USD prices are multiplied by 1,000,000 for models.dev's USD per-1M-token cost fields, including input, output, reasoning, cache-read, and cache-write prices. Non-USD prices are omitted rather than mislabeled as USD.
• Canonical mapping: canonical_slug is preferred, then the namespaced author/id route is used as a fallback. Known namespace differences (zhipu → zhipuai, moonshot → moonshotai, and mistral → mistral) are resolved through models.dev metadata. Existing base_model / base_model_omit factoring is preserved.
• Capabilities: attachment/modalities come from architecture; tool calling, structured output, temperature, and fallback reasoning detection come from supported_parameters; exact reasoning effort/budget controls come from the catalog's reasoning object. Missing or all-false reasoning controls are treated as unknown and inherit canonical or authored truth.
• Lifecycle safety: release dates come only from the catalog's explicit lifecycle fields or canonical metadata; the catalog insertion timestamp is never treated as a model release date. All 122 live routes now expose authoritative release dates and are included.
• Preservation: authored descriptions, lifecycle dates, knowledge cutoffs, status, interleaving, input limits, pricing tiers, and reasoning options survive when the source is not authoritative.
• Determinism: a clean sync created 122 TOMLs; the immediately following dry run reported 0 created, 0 updated, 0 removed, 122 unchanged.

Validation

• bun test packages/core/test/sync.test.ts — 57 passed
• bun --cwd packages/sdk test — 23 passed
• bun models:sync eurouter
• bun models:sync eurouter --dry-run — clean second run
• bun validate
• git diff --check

The root bun test command also runs a repository metadata audit that currently fails on dev for 30 pre-existing open-weight metadata entries without weights links; the same failure was reproduced in an untouched origin/dev worktree. The prescribed SDK test command above generates its ignored snapshot first and passes.

E2E model coverage

EUrouter's existing route E2E fixtures exercise openai/gpt-4o-mini for Chat Completions / Responses and mistral/mistral-embed for Embeddings. Adapter regression fixtures additionally cover claude-sonnet-5 and gpt-5.1. This data-only contribution did not issue paid inference requests.

Sources

• EUrouter model catalog API
• Models API documentation
• Quickstart and OpenAI-compatible base URL
• Model directory and per-1M pricing
• Official EUrouter logo asset

[DavidLoDico]

Adding EUrouter, the european AI gateway. Use any AI model, keep your data in the EU.

[github-actions]

Action items

• [medium] [possible mistake] providers/eurouter/models/qwen3-235b-a22b-thinking.toml:6, qwen3-30b-a3b-thinking.toml:6, qwen3-vl-235b-a22b-thinking.toml:6, deepseek-r1-distill-llama-70b.toml:7, green-r.toml:7, green-r-raw.toml:7 - Check: reasoning = false on models whose names or descriptions explicitly indicate reasoning/thinking capability. Why: These inline models have reasoning = false despite names containing "thinking" (e.g. "Qwen3 235B A22B Thinking") or descriptions mentioning "chain-of-thought reasoning with `` tokens" (deepseek-r1-distill-llama-70b) or "Reasoning model" (green-r, green-r-raw). The sync sets reasoning = false as a fallback when the EUrouter API omits the `reasoning` object and no matching `models/` metadata entry exists for `base_model` inheritance. In contrast, `qwen3-next-80b-a3b-thinking.toml` correctly inherits `reasoning = true` via `base_model = "alibaba/qwen3-next-80b-a3b-thinking"` with `reasoning_options = []`, proving that "thinking" variants on EUrouter can and should be reasoning models. Users filtering for reasoning models on models.dev would miss these models. Action: Verify whether the EUrouter API actually omits reasoning metadata for these specific models. If reasoning is exposed but unreported, manually set `reasoning = true` and `reasoning_options = []` in each affected TOML — the sync preserves existing `reasoning` and `reasoning_options` values on subsequent runs via the `existing?.reasoning` and `existing?.reasoning_options` fallbacks in `buildEUrouterModel`.

[DavidLoDico]

I'll review the action items. Update coming soon.

[github-actions]

No actionable findings.

[DavidLoDico]

@rekram1-node hey, wondering if you need any more info to merge this PR?