Add Omnious as an inference provider

[LJ-XRPL]

Omnious serves an OpenAI-compatible endpoint over a live per-request auction: providers bid to serve each call, and the winner is paid a second-score price capped by a genuine rival. A class's served price therefore moves with the book rather than sitting on a rate card.

That makes a hand-authored [cost] stale on arrival, so this adds a sync module rather than static pricing. GET /v1/models is public (no key required) and reports the current best bid per class as best_in / best_out in USDC base units per 1M tokens; the module converts those to USD and writes them as [cost]. It runs in the aggregators group so the existing automation keeps it current.

What each entry carries

Price is the only provider-specific fact Omnious publishes — there's no context, limit or capability data in the catalog. So each entry is base_model plus [cost] and inherits everything else, which means it can't contradict the base model:

base_model = "zhipuai/glm-5.2"

[cost]
input = 0.64526
output = 2.027962

Attribution

Classes are attributed by the router's issuer field, mapped to a models.dev author. When the router reports no issuer, the module falls back to a suffix match and takes it only when exactly one author publishes that model ID, so an ambiguous short name is skipped rather than assigned to whichever author sorted first.

A class is skipped when models.dev doesn't carry its base model yet, or when that base model has no limit.output to inherit (sakana/fugu-ultra, thinkingmachines/inkling) — inventing a ceiling seemed worse than omitting the row. 128 of 291 classes map today; the rest need their author metadata added under models/ first.

Lookups resolve through an index built from models/, keyed lowercase and mapped back to the ID as authored, because Omnious class names are lowercase while some base models aren't (minimax/MiniMax-M2). That also avoids an existsSync probe answering yes on a case-insensitive volume and then failing catalog generation on Linux.

Verification

• bun models:sync omnious → 128 created
• bun validate → passes
• Spot-checked generated [cost] against the live endpoint (glm-5.2 → 0.64526 / 2.027962, kimi-k3 → 3.3 / 16.5)

[github-actions]

Action items

• [critical] [violation] providers/omnious/logo.svg:1 - Check: New provider logo.svg must use currentColor, with no fixed size or hardcoded colors. Why: The logo sets width/height and hardcodes strokes/fills (#6c6a64, #a3a19a, #e6e4df, #a98e6f), so it will not theme correctly and fails the new-provider logo blocker. Action: Rewrite the SVG with a square viewBox, drop fixed dimensions, and express all fills/strokes as currentColor (use opacity if multi-tone is needed).
• [high] [violation] providers/omnious/models/gpt-5.4.toml:2 - Check: Multi-model relays with reasoning = true must author host controls from the lab + same-surface peers; [] means no caller control, not uncertainty. Why: Omnious is an OpenAI-compatible multi-lab auction gateway, yet ~100 reasoners ship reasoning_options = [] (e.g. gpt-5.4, deepseek-v4-pro, claude-sonnet-5, Gemini/Qwen/GLM families). Lab/peers expose real controls (GPT effort L–xhigh; DeepSeek V4 toggle+high/max; Claude Sonnet 5 effort; etc.). The sync module only emits base_model+cost, so the runner will keep injecting [] for new reasoners. Action: For each reasoner, set options from first-party + OpenAI-compat peers (or document affirmative evidence Omnious exposes no control). Update translateModel / hand-authored files so automation preserves real options instead of defaulting to [].
• [high] [violation] packages/core/src/sync/providers/omnious.ts:487 - Check: Sync must not delete catalog rows when a source model is only temporarily unusable (sync.md: return undefined only when absence should remove the local file). Why: translateModel returns undefined when best_in/best_out is missing or <= 0, and parseModels drops rows that fail schema. With default deleteMissing, the next hourly run removes the TOML even if the class is still served—especially risky for a live auction book. Action: When a class is attributable but has no live bid, preserve the existing entry (update cost only when both sides are priced), and/or set deleteMissing: false for price-only skips; reserve hard deletes for classes truly gone from the catalog.

[github-actions]

Action items

• [high] [violation] providers/omnious/logo.svg:1 - Check: New-provider logo must use currentColor, no fixed size, no hardcoded colors. Why: The SVG sets width="144" height="144" and hardcodes stroke/fill colors (#6c6a64, #a3a19a, #e6e4df, #a98e6f), so it will not theme with the site and fails the new-provider logo blocker. Action: Rewrite logo.svg with a square viewBox, no width/height attributes, and currentColor for fills/strokes.
• [high] [violation] providers/omnious/models/gpt-5.4.toml:2 - Check: Multi-model relays must set reasoning_options from the lab entry + same-surface peers; [] means no caller control, not uncertainty. Why: Omnious is an OpenAI-compatible aggregator, yet every reasoning model is authored with reasoning_options = [] (including GPT-5.x, Claude Sonnet/Fable 5, DeepSeek V4, Gemini, Qwen, Kimi, MiniMax, etc.). Lab/peers expose real controls (e.g. OpenAI/OpenRouter/Vercel/LLM Gateway GPT-5.4: effort none|low|medium|high|xhigh; DeepSeek V4: toggle + high/max; Claude Sonnet 5 peers: effort levels). Shipping blank options misrepresents the API surface for ~100 reasoners. Action: For each reasoning = true Omnious model, copy the lab + openai-compatible peer option set (effort/toggle/budget as applicable); keep [] only where lab/peers affirm always-on/no control (e.g. classic R1-style). After authoring, later cost-only syncs will preserve those options via the runner.