feat: add Cafe24 LLM Router provider

[jiho09]

What

Adds Cafe24 LLM Router (https://llm-router.cafe24.com) as a provider — an OpenAI-compatible LLM gateway by Cafe24, a KRX-listed Korean e-commerce & hosting platform company. We are the Cafe24 LLM Router team and will maintain this entry.

Contents

• providers/cafe24/provider.toml — @ai-sdk/openai-compatible, api https://llm-router.cafe24.com/api/v1, env CAFE24_API_KEY
• 8 relayed models via base_model per AGENTS.md (Qwen3 32B, MiniMax M3, GPT OSS 20B/120B, Claude Haiku 4.5 / Sonnet 5 / Opus 5 / Fable 5) — provider files are override-only: USD costs (converted from our published KRW prices) and [limit] overrides taken from our live GET /api/v1/models metadata
• cafe24/auto — host-unique auto-routing model, inline definition with no [cost] (variable pricing), following the openrouter/auto precedent
• reasoning_options = [] on relayed models — our gateway does not currently document reasoning-control parameters, so no controls are claimed
• Monochrome currentColor logo — vector-traced from the official cafe24 wordmark

Validation

bun validate passes (exit 0) with this change.

Note: full catalog listing is 100+ models — this initial entry curates flagship models whose lab metadata already exists; we plan follow-ups for the rest.

Evidence

• Pricing: KRW list prices from https://llm-router.cafe24.com/models (captured 2026-08-04). The site prices at a fixed ₩1,430/USD rate — corroborated by Claude list prices dividing exactly to the labs' published USD prices (Haiku 4.5 ₩1,430/₩7,150 → $1/$5; Sonnet 5 ₩2,860/₩14,300 → $2/$10). Each priced file carries this provenance in a leading comment.
• Limits: taken from the live GET /api/v1/models metadata (max_context_tokens / max_output_tokens), overriding the lab value only where the gateway serves less.
• Reasoning controls: verified per-model against the live gateway on 2026-08-04 — GPT-OSS forwards reasoning_effort (high=76 vs low=22 completion tokens), Qwen3-32B forwards the thinking toggle (320 vs 2 tokens), Claude models and MiniMax M3 accept but do not forward either OpenAI-style or Anthropic-style reasoning parameters (no reasoning-token accounting, no differential).

[github-actions]

Action items

• [high] [violation] providers/cafe24/models/anthropic/claude-opus-5.toml:2 - Check: Relay reasoning_options must match lab/same-surface peer controls; [] only for affirmative no caller control, not uncertainty. Why: Cafe24 is a multi-model OpenAI-compatible gateway. All eight relayed models inherit reasoning = true and set reasoning_options = [], while the PR only says controls are “not currently document[ed].” Lab/peers expose real controls for most of these: Claude Opus/Sonnet/Fable 5 → effort (low/medium/high/xhigh/max on Anthropic/OpenRouter); Haiku 4.5 → budget_tokens (lab) / toggle+budget (OpenRouter); gpt-oss-20b/120b → effort low/medium/high (OpenRouter/Groq/Kilo); Qwen3-32B → toggle+budget (Alibaba); MiniMax-M3 → toggle (MiniMax). Empty options here misrepresent API capability if fields are passed through. Action: For each model, either (a) author the host’s real controls from lab + peer baselines (e.g. Claude effort; gpt-oss L/M/H; Qwen toggle/budget; MiniMax toggle), or (b) keep [] only with an affirmative top-of-file note that this gateway strips/ignores reasoning controls—not that docs are missing.
• [medium] [violation] providers/cafe24/models/MiniMaxAI/MiniMax-M3.toml:4 - Check: Non-USD costs must be converted to USD/MTok with rate/date in a leading TOML comment. Why: The PR states costs were converted from published KRW prices, but none of the cost-bearing model files include a top-of-file FX rate/date comment (required by AGENTS.md; see GreenPT EUR examples). Without that, reviewers cannot audit the USD figures. Action: Add a leading comment on each priced model (or a shared header pattern) with KRW source prices, USD/KRW rate, and capture date; or state that the API already bills in USD and drop the KRW-conversion claim.
• [low] [possible mistake] providers/cafe24/provider.toml:5 - Check: Data-changing PRs should cite first-party pricing/docs/API evidence mapped to claims. Why: Costs and limit overrides are material, but the PR body only points at the product site/docs generally and “live GET /api/v1/models” without URLs or a claim→source map (KRW list prices, conversion, per-model context/output caps). Action: In the PR body, link Cafe24 pricing and models API/docs and note which claim each source supports (per-model USD/MTok, limit overrides).

[github-actions]

Action items

• [high] [violation] providers/cafe24/models/anthropic/claude-sonnet-5.toml:2 - Check: Relay reasoning_options must match lab/same-surface peer controls; [] only for affirmative no caller control. Why: Cafe24 is a multi-model OpenAI-compatible gateway. All eight relayed models set reasoning_options = [] while lab/peers expose controls (Claude Sonnet/Opus/Fable 5 → effort low/medium/high/xhigh/max; Haiku 4.5 → toggle + budget; GPT-OSS → low/medium/high; Qwen3-32B → toggle + budget; MiniMax-M3 → toggle). The PR only says controls are undocumented, which is uncertainty—not an affirmative “gateway strips/ignores reasoning params.” Action: For each model, author the host’s real wire controls (or, if none are forwarded, state that affirmatively in a leading comment and keep [] only then). Do not use [] solely because docs omit the params.
• [medium] [violation] providers/cafe24/models/anthropic/claude-sonnet-5.toml:4 - Check: Non-USD costs must be converted to USD/MTok with rate/date in a top-of-file comment. Why: The PR body says prices were converted from published KRW, but none of the cost-bearing model files include a leading KRW→USD rate/date (or source) comment. Same gap on MiniMax-M3, Qwen3-32B, Claude Haiku/Opus/Fable, and both GPT-OSS entries. Action: Add a leading comment on each priced file (or shared header pattern) with KRW list price, FX rate, capture date, and resulting USD/MTok values.
• [low] [possible mistake] providers/cafe24/provider.toml:5 - Check: Data-changing PRs should cite pricing/docs/API evidence mapped to claims. Why: Body mentions the docs URL and live /api/v1/models for limits, but does not cite a pricing page, KRW list prices, or FX source for the published costs/limits. Action: Add first-party pricing + FX citations in the PR body, noting which models/fields each supports.

[jiho09]

Addressed all three action items in 8deaa41:

[high] reasoning_options — replaced the blanket [] with per-model values from live-gateway probes run 2026-08-04 (identical prompts, real calls):

• GPT-OSS 20B/120B: reasoning_effort is forwarded — completion tokens scale with effort (high=76 vs low=22) → effort: low/medium/high
• Qwen3-32B: thinking toggle is forwarded via chat_template_kwargs.enable_thinking (default/thinking=320 tokens vs disabled=2 tokens) → toggle
• Claude Haiku 4.5 / Sonnet 5 / Opus 5 / Fable 5 and MiniMax M3: OpenAI-style reasoning_effort and Anthropic-style thinking are accepted but not forwarded — no reasoning-token accounting in usage, no output differential (e.g. Sonnet 5: 11 vs 13 completion tokens with/without effort) → affirmative [], stated in a leading comment on each file

[medium] cost provenance — every priced file now opens with the KRW list price, source (https://llm-router.cafe24.com/models), capture date (2026-08-04), and the conversion at the site's fixed ₩1,430/USD rate. The fixed rate is externally corroborated: the site's Claude list prices divide exactly to the lab's published USD prices (e.g. Haiku 4.5 ₩1,430/₩7,150 → $1/$5).

[low] PR-body citations — pricing/limits/FX evidence added to the description.

bun validate passes.

[github-actions]

Action items

• [medium] [possible mistake] providers/cafe24/models/MiniMaxAI/MiniMax-M3.toml:8 - Check: Relay reasoning_options must reflect this host’s real control surface (lab/peer baseline + what the gateway actually forwards), not untested param families. Why: This file sets reasoning_options = [] after probing only OpenAI reasoning_effort and Anthropic thinking, but the same Cafe24 gateway is already documented (on Qwen3-32B) as forwarding chat_template_kwargs. Open-weight MiniMax-M3 peers expose thinking via chat_template_kwargs.thinking_mode (enabled/disabled/adaptive), and first-party MiniMax authors a toggle — so [] may under-report a real control. Action: Probe chat_template_kwargs.thinking_mode (and any MiniMax-native toggle this host passes through). If it changes reasoning behavior, set reasoning_options = [{ type = "toggle" }] with a leading wire-path comment; only keep [] if that surface is also a no-op.

[jiho09]

Good catch — probed chat_template_kwargs.thinking_mode on MiniMax-M3 as suggested (live gateway, 2026-08-04): it is forwarded — identical prompt yields 26 (enabled) / 2 (disabled) / 16 (adaptive) completion tokens. Corrected reasoning_options from [] to [{ type = "toggle" }] with a leading wire-path comment in the latest commit. bun validate passes.

[github-actions]

No actionable findings.