feat: add Melious provider

[sahil-melious]

Adds Melious — a European OpenAI-compatible inference host
that serves open-weight models only — as a provider.

providers/melious/: provider.toml, logo.svg, and 53 chat models.

Data provenance

• Models, capabilities, context windows: GET https://api.melious.ai/v1/models?include_meta=true
• Prices: https://melious.ai/hub/models — listed in EUR, converted at 1.17 USD/EUR
  (ECB reference rate, captured 2026-08-20). Each file's leading comment records the
  rate, the date and the source.
• Provider docs: https://melious.ai/docs/reference/models

Model files

48 of 53 use base_model against existing models/ lab metadata and are
override-only — cost, reasoning_options, interleaved, plus real deltas such as
a narrower limit.context.

Five have no lab entry upstream and are defined inline: gemma-3-27b-it,
holo2-30b-a3b, llama-3.1-405b-instruct, hermes-4-405b, hermes-4-70b. Every
other host of those same models does the same (greenpt, stackit, nebius,
cortecs), and their names and descriptions reuse the strings already normalised
across those entries.

Reasoning

reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }] on every
reasoner — that is this host's documented control
(docs). No toggle: an explicit
reasoning_effort: "none" still returns reasoning, so there is no reachable off state.

interleaved.field = "reasoning_content" was set per model from live responses, not
assumed. Both Hermes 4 models are classified as reasoners by the catalog but inline
their thinking into content, so they deliberately carry no interleaved. Conversely
devstral-2-123b-instruct-2512, gemma-3-27b-it and holo2-30b-a3b are classified
non_reasoning yet return chain-of-thought in reasoning_content, so they are
recorded as reasoners; each file notes the discrepancy.

Limits

limit.output is measured, not inherited: the catalog reports max_output_tokens as
null for 51 of 53 models. Values come from descending a ladder from the context
window and bisecting to ±1024. Where the host accepts max_tokens equal to the full
window, limit.output is the window; where output shares a smaller window with the
prompt, that window is published; each file says which case applies.

Verification

• bun validate — passes
• bun run test in packages/sdk — passes
• OpenCode, per the README's manual-testing flow:

 cd packages/web && bun run build
 OPENCODE_MODELS_PATH="dist/_api.json" MELIOUS_API_KEY=… opencode 

opencode auth list picks the key up from MELIOUS_API_KEY, opencode models lists
all 53 as melious/*, and opencode run completes with tool calls and reasoning
parts against a real key with no opencode.json present.

Checklist

• Compliant logo.svg — square viewBox, currentColor, no fixed width/height
• Non-lab hosts use base_model; inline entries follow existing precedent
• base_model files are override-only, no restated identical fields
• reasoning = true ⇒ reasoning_options set
• Costs are USD/MTok, with conversion rate and date in a top-of-file comment
• bun validate passes

[github-actions]

Action items

• [high] [violation] providers/melious/models/gemma-3-27b-it.toml:13 - Check: Non-lab hosts must use base_model; missing lab metadata must be added under models/. Why: Melious did not create these models, and each has a nameable lab identity (google/gemma-3-27b-it, H Company Holo2, meta/llama-3.1-405b-instruct, NousResearch Hermes 4 405B/70B). Full inline third-party definitions are only allowed for unique-to-host/private aliases; peer hosts defining them inline does not create an exception. Action: Add complete lab files under models/<lab>/… for all five, then rewrite the Melious entries as override-only base_model files (cost, reasoning_options, interleaved, and real host deltas only). Same applies to holo2-30b-a3b.toml, llama-3.1-405b-instruct.toml, hermes-4-405b.toml, and hermes-4-70b.toml.
• [high] [possible mistake] providers/melious/models/qwen3.5-122b-a10b.toml:18 - Check: Cache pricing should be a discounted read rate relative to input. Why: cache_read = 0.1404 is higher than input = 0.117 (about 120% of input). Other Melious cache rows are ~20–40% of input; 0.1404 is also identical to kimi-k2.5’s cache_read, which looks like a copy-paste error. Action: Re-check Melious EUR cache pricing for this model, reconvert at 1.17 USD/EUR, and correct cache_read (or remove it if cache is not billed).