feat: add Lyceum provider

[Casparlehmkuehler]

Lyceum (https://lyceum.technology) is an EU-hosted, OpenAI-compatible inference API.
42 model ids at https://api.lyceum.technology/openai/v1, auth LYCEUM_API_KEY.

Lyceum hosts other labs' models, so every provider entry is base_model +
overrides. Six lab entries did not exist yet and are added under models/:
nousresearch/hermes-4-{70b,405b}, openbmb/minicpm-v-4_5,
google/gemma-3-27b-it, nvidia/cosmos3-super-reasoner,
alibaba/qwen3-30b-a3b-instruct-2507, alibaba/qwen3-embedding-8b.

Every provider-side value was measured against the live API on 2026-09-02, not
assumed:

• cost: GET /api/v2/external/pricing?resource=serverless_inference, USD per
  token converted to USD/MTok.
• reasoning_options: each reasoning_effort value posted to
  /chat/completions (3 attempts, to separate real rejections from transients).
  none is listed only where it demonstrably suppresses reasoning. [] means
  the host accepts no caller control, not that it went untested.
• interleaved: which response key carries the reasoning text. true where it
  is reasoning, which the schema's field enum does not cover.
• attachment: an image_url content part posted per model; text-only models
  reject it, so the signal discriminates.
• limit.output: the proxy rejects max_tokens above 65536 for every model, so
  entries resolve to at most that. Three models cap lower and were bisected:
  openbmb/minicpm-v-4_5 24576, qwen/qwen2.5-vl-72b-instruct and
  qwen/qwen3-32b sit under the global cap already.

Two id families are Lyceum-specific aliases, both resolved server-side and
priced identically to their target: lyceum/{simple,complex,reasoning} (router
aliases) and the -instant suffix (pins enable_thinking = false for clients
that cannot set request parameters).

bun validate passes.

Happy to supply a test key to a maintainer out of band if CI needs one.

[github-actions]

Action items

• [high] [violation] providers/lyceum/models/deepseek/deepseek-v4-pro.toml:2 - Check: Reasoning options must follow lab/peer baseline for this host, not a full schema enum dump. Why: Lyceum is a multi-model OpenAI-compatible relay. First-party DeepSeek V4 Pro is toggle + high/max (Flash is toggle + low/high/max); established relays stay near that narrow set. This entry (and deepseek-v4-flash-0731.toml) authors the full effort enum none|minimal|low|medium|high|xhigh|max, which AGENTS.md / the audit skill forbid as invented universal levels—API acceptance alone is not distinct effort. Action: Replace with the host-true subset of DeepSeek’s controls (typically none in effort or a documented toggle, plus only levels that change behavior on Lyceum, e.g. high/max or Flash’s low/high/max), with a leading wire comment for the exact field.
• [high] [violation] providers/lyceum/models/z-ai/glm-5.2.toml:2 - Check: GLM reasoning_options must match lab/same-surface peers, not a full enum. Why: First-party Z.ai GLM-5.2 is effort high/max (with none/minimal as skip); GLM-5.3 / 5.3-Flash are low/high/max. Lyceum dumps the full enum on glm-5.2, glm-5.3, glm-5.3-flash, and the lyceum/reasoning alias, and GLM-5.1 uses a graded list while the lab is toggle-only. That mis-documents caller controls vs the underlying models. Action: Author only levels Lyceum actually exposes with distinct effect, aligned to each GLM generation (5.1 toggle or equivalent; 5.2 high/max ± real off; 5.3/Flash low/high/max ± real off), and mirror the same set on router aliases.
• [high] [violation] providers/lyceum/models/moonshotai/kimi-k2.6.toml:2 - Check: Kimi reasoning_options baseline is lab + peers, not a full dump. Why: Moonshot K2.6 is toggle-only; K3 is toggle + low/high/max. Lyceum sets K2.6 to the full effort enum and K3 to only none/max, which neither matches the lab surface nor a documented peer intersection. Action: Set K2.6 to the control Lyceum actually forwards (toggle and/or the real effort values), and set K3 to the measured subset of low/high/max plus off via none or toggle—not a full enum or an unexplained none/max-only pair.
• [high] [violation] providers/lyceum/models/qwen/qwen3.5-9b.toml:2 - Check: Qwen3.5 relay options should follow Alibaba/peer controls for this API shape. Why: First-party Qwen3.5 is toggle (+ often budget_tokens on Alibaba chat), not a seven-value effort ladder. Lyceum applies the full enum on qwen3.5-9b, qwen3.5-397b-a17b, and lyceum/simple. That invents GPT-style graded effort for a toggle/budget family. Action: Use the controls Lyceum really implements for these IDs (e.g. reasoning_effort values that map to thinking on/off and any real grades, or toggle + budget if that path exists)—not the full schema enum—and keep router aliases in sync.
• [medium] [violation] providers/lyceum/models/nvidia/nemotron-3-super-120b-a12b.toml:2 - Check: Nemotron options must not invent graded effort when lab/peers are toggle/budget. Why: NVIDIA first-party Super/Ultra/Nano are toggle (Omni adds budget_tokens). Lyceum sets Super to low/medium/high, Ultra and Omni to the full effort enum, and Nano to [] while still marking interleaved reasoning. That conflicts with the lab baseline and the “no full enum / no invented L/M/H” rule unless each level is proven distinct on this host. Action: Align each Nemotron entry to measured Lyceum controls without dumping the schema enum; prefer lab-like toggle (and Omni budget only if present), or a proven effort subset—not Ultra/Omni full dumps or Super L/M/H by default.
• [medium] [possible mistake] providers/lyceum/models/z-ai/glm-5.1.toml:2 - Check: Provider overrides of lab modalities/attachment must be real host deltas. Why: Lab zhipuai/glm-5.1 is text-only (attachment = false, input ["text"]). This file forces attachment = true and input = ["text", "image"]. That is a large capability claim beyond the shared lab identity; a single image_url accept is weak evidence that the model is multimodal. Action: Confirm vision is truly supported end-to-end on Lyceum for GLM-5.1; if not, drop the attachment/modalities overrides and keep lab text-only metadata.
• [medium] [possible mistake] providers/lyceum/models/nvidia/llama-3_1-nemotron-ultra-253b-v1.toml:2 - Check: reasoning = false overrides on a reasoning lab model need host-specific evidence. Why: nvidia/llama-3.1-nemotron-ultra-253b is reasoning = true. Overriding to false without a leading rationale (unlike MiniCPM / -instant aliases) risks cataloging a reasoner as non-reasoning if the host still emits CoT. Action: Document/verify that this Lyceum ID never returns reasoning; if it does, keep reasoning = true and set accurate reasoning_options / interleaved.
• [low] [possible mistake] models/nousresearch/hermes-4-405b.toml:16 - Check: New lab entries should include complete limit defaults when known. Why: Hermes 4 and Cosmos3 Super Reasoner lab files omit limit.output (providers only patch output). That leaves incomplete lab metadata for other hosts and makes inheritance easy to get wrong. Action: Add authoritative lab limit.output (and any other known lab defaults) from first-party/model cards where available, keeping Lyceum’s 65_536 as a provider override only when the host is lower.

[Casparlehmkuehler]

Thanks, all eight items addressed in the amended commit.

The full-enum lists were the wrong call: I had measured which reasoning_effort
values the API accepts, and you are right that acceptance is not a distinct
level. Re-measured with 5 attempts per value (the first pass had transients that
made a couple of models look narrower than they are) and authored the
intersection of the first-party entry and what this host really exposes:

• deepseek-v4-pro toggle + high/max; deepseek-v4-flash-0731 toggle +
  low/high/max, matching providers/deepseek.
• glm-5.1 toggle; glm-5.2 toggle + high/max; glm-5.3 and
  glm-5.3-flash toggle + low/high/max, matching providers/zhipuai. The
  toggle is real on top of the graded set: none suppresses the reasoning,
  which is also what the -instant aliases pin.
• kimi-k2.6 toggle; kimi-k3 toggle + low/high/max. medium is the
  only value K3 rejects, so it now matches providers/moonshotai exactly.
• qwen3.5-9b and qwen3.5-397b-a17b toggle. No budget field is reachable
  here, so unlike Alibaba's own entries there is no budget_tokens.
• qwen3.8-27b and qwen3.8-flash-next toggle + low/medium/xhigh, which
  is providers/alibaba/models/qwen3.8-flash.toml's effort set; high is
  rejected upstream. qwen3.8-2.4t-a95b gets the same effort set with no
  toggle, since it rejects none.
• Nemotron: Omni none/low/medium/high/max and Ultra
  none/medium/high, following the relay peers. Super keeps
  low/medium/high because this host rejects none, so there is genuinely
  no off switch; Nano-30B keeps [] because none is accepted but does not
  suppress anything. Both now say so in a leading comment.

Every toggle and effort entry now carries a leading comment naming the wire
control, per AGENTS.md.

On the two possible mistakes:

• glm-5.1 vision: you were right to doubt it. Sent a solid-colour test image
  and asked for the colour; GLM-5.1 answers "White" for a red image, so it
  accepts the part without reading it. Dropped the attachment and modalities
  overrides and kept the lab's text-only metadata. Re-ran the same check on the
  entries that do claim vision beyond their lab row: qwen2.5-vl-72b-instruct
  and minicpm-v-4_5 name the colour correctly, so those overrides stand.
• llama-3_1-nemotron-ultra-253b-v1: also right. With detailed thinking on as
  the system prompt it does reason, into a <think> block inside content.
  Removed the reasoning = false override; it is now reasoning = true with
  reasoning_options = [] and a comment explaining that the switch is the
  system prompt, not a request field. The same test caught qwen3-32b, which
  reasons via Qwen's /think convention, so it changed the same way. Neither
  sets interleaved, since the block is in content rather than a side channel.

Lab limit.output added: Hermes 4 70B/405B 131_072, Cosmos3 Super Reasoner
256_000, MiniCPM-V 4.5 32_768. Lyceum's own ceilings stay as provider overrides
(65_536 platform-wide, 24_576 for MiniCPM).

bun validate passes.

[github-actions]

Action items

• [high] [violation] providers/lyceum/models/deepseek/deepseek-v4-flash-0731.toml:4 - Check: When off is reasoning_effort = none among graded levels, author effort with none in values and no toggle (AGENTS.md → Reasoning options; same-surface peers such as NVIDIA NIM / Baseten / Above / Kosmik / Groq). Why: These files document off as reasoning_effort = none on the same field as graded effort, then encode toggle + effort without none. That is not a separate on/off control; clients will not learn to send none, and the shape conflicts with established OpenAI-compatible peers. Action: Drop toggle and put none in effort.values. Expected sets from the file comments: deepseek-v4-flash-0731 → none|low|high|max; deepseek-v4-pro, lyceum/reasoning, z-ai/glm-5.2 → none|high|max; moonshotai/kimi-k3 → none|low|high|max; qwen/qwen3.8-27b, qwen/qwen3.8-flash-next → none|low|medium|xhigh; z-ai/glm-5.3, z-ai/glm-5.3-flash → none|low|high|max. Keep a leading wire comment like # Effort: reasoning_effort = none|…. Apply the same fix in: providers/lyceum/models/deepseek/deepseek-v4-pro.toml, providers/lyceum/models/lyceum/reasoning.toml, providers/lyceum/models/moonshotai/kimi-k3.toml, providers/lyceum/models/qwen/qwen3.8-27b.toml, providers/lyceum/models/qwen/qwen3.8-flash-next.toml, providers/lyceum/models/z-ai/glm-5.2.toml, providers/lyceum/models/z-ai/glm-5.3.toml, providers/lyceum/models/z-ai/glm-5.3-flash.toml.

[Casparlehmkuehler]

Fixed. toggle is gone from every entry whose off switch is
reasoning_effort = none on the same field as the graded levels; those are now
a single effort list with none in values, exactly the sets you listed:

• deepseek-v4-flash-0731, kimi-k3, glm-5.3, glm-5.3-flash → none|low|high|max
• deepseek-v4-pro, glm-5.2, lyceum/reasoning → none|high|max
• qwen3.8-27b, qwen3.8-flash-next → none|low|medium|xhigh

The leading comments now say which way the unlisted values go, since it differs
by host: kimi-k3 rejects medium, the qwen3.8 pair rejects high and max,
and the GLM and DeepSeek entries accept everything but collapse it onto the
listed levels the way their first-party APIs document.

toggle survives only on the five ids with no graded levels at all
(kimi-k2.6, glm-5.1, qwen3.5-9b, qwen3.5-397b-a17b, lyceum/simple),
which matches their lab entries.

bun validate passes.

[github-actions]

Action items

• [high] [violation] providers/lyceum/models/meta-llama/llama-3.3-70b-instruct.toml:3 - Check: Override-only base_model files must not restate identical base fields. Why: models/meta/llama-3.3-70b-instruct.toml already has attachment = false and text-only modalities, so this restates an unchanged value. Action: Remove the redundant attachment = false (and keep only real host deltas such as cost).
• [medium] [possible mistake] providers/lyceum/models/z-ai/glm-5.3.toml:4 - Check: Relay reasoning_options must match this host’s real controls and must not invent levels beyond lab/peer baseline without verified effect. Why: First-party providers/zai/models/glm-5.3.toml and established peers author low|high|max and document that GLM-5.3 always reasons (thinking cannot be disabled). This entry adds none as an off switch, which contradicts that baseline. Action: Re-verify on Lyceum whether reasoning_effort=none actually suppresses reasoning end-to-end; if not, drop none and use ["low", "high", "max"] (and apply the same correction to providers/lyceum/models/z-ai/glm-5.3-instant.toml only if that alias is meant to stay non-reasoning via name encoding, not via effort).
• [medium] [possible mistake] providers/lyceum/models/qwen/qwen3.8-flash-next.toml:4 - Check: Effort sets for always-on lab models should not list none unless this host truly turns reasoning off. Why: Lab metadata for alibaba/qwen3.8-flash-next states thinking is always on with low|medium|xhigh, matching peers such as Cortecs; this file still includes none as a disable value. Action: Confirm reasoning_effort=none actually disables reasoning on Lyceum; if it is rejected or ignored, remove none and keep ["low", "medium", "xhigh"].
• [medium] [possible mistake] providers/lyceum/models/minimax/minimax-m3.toml:3 - Check: On relays, reasoning_options = [] means affirmative no caller control, not “only reasoning_effort was tried.” Why: First-party MiniMax M3 and several same-surface peers expose a real on/off control (toggle), while this entry sets [] after testing only reasoning_effort. The same pattern appears on providers/lyceum/models/lyceum/complex.toml (alias of M3). Action: Verify whether Lyceum forwards any non-reasoning_effort disable/enable control used by MiniMax/peers; if yes, author that control (with a leading wire comment), otherwise keep [] and note the affirmative no-control finding in the leading comment.
• [medium] [possible mistake] providers/lyceum/models/nvidia/nvidia-nemotron-3-nano-30b-a3b.toml:3 - Check: [] on a relay of a lab model that has caller controls needs affirmative host evidence. Why: First-party providers/nvidia/models/nvidia/nemotron-3-nano-30b-a3b.toml authors toggle, while this file sets [] because reasoning_effort=none does not suppress reasoning. That may miss a different wire control the host still accepts. Action: Confirm whether Lyceum exposes any on/off control for this model beyond reasoning_effort; if a real toggle exists, author it with a leading wire comment, otherwise keep [] with an explicit no-control note.

[Casparlehmkuehler]

Two of the four doubts were real bugs, one was a good question with a
measurable answer, and one rests on a lab row that says the opposite.

Fixed: nemotron-3-nano-30b-a3b does have a toggle. You were right that
[] after testing only reasoning_effort was not affirmative. The control is
chat_template_kwargs.enable_thinking = false, and it is a real disable, not
response filtering: on a fixed prompt the completion drops from 278 to 131
tokens and the reasoning field disappears. Now [{ type = "toggle" }] with
that field named in the leading comment.

Checked and kept [], now stated affirmatively. For minimax-m3,
minimax-m2.5, kimi-k2.7-code, qwen3-next-80b-a3b-thinking and the
lyceum/complex alias I tried reasoning_effort, enable_thinking,
chat_template_kwargs.enable_thinking, thinking, thinking.type and
reasoning.enabled; every one leaves the reasoning on. Each file's leading
comment now lists what was tried.

One trap worth recording: on minimax-m2.5, reasoning.exclude = true makes
the reasoning vanish from the response, which looks like a toggle. It is not.
With max_tokens = 600 the completion still bills all 600 tokens and returns
empty content, so the model reasoned and the output was discarded. Not authored
as a control.

none genuinely disables on the GLM, DeepSeek, Kimi and qwen3.8 entries.
Re-verified end to end with a 400-token budget and a reasoning-inducing prompt,
three repeats each. Without the flag every one returns a populated reasoning
channel; with reasoning_effort = none all three repeats come back with no
reasoning channel and no <think> block:

model	no flag	none x3
glm-5.1	reasoning_content	off, off, off
glm-5.2	reasoning	off, off, off
glm-5.3	reasoning	off, off, off
glm-5.3-flash	reasoning_content	off, off, off
qwen3.8-27b	reasoning_content	off, off, off
qwen3.8-flash-next	reasoning_content	off, off, off
deepseek-v4-pro	reasoning	off, off, off
deepseek-v4-flash-0731	reasoning	off, off, off
kimi-k3	reasoning	off, off, off

So none stays in those values. Lyceum serves GLM-5.3 and the qwen3.8 pair
through an upstream that honours the switch even where the lab row says thinking
is always on; if you would rather the catalog not diverge from the lab baseline
on that point, say so and I will drop none and lose the off switch in the
metadata.

llama-3.3-70b-instruct: the attachment = false override is a real delta.
models/meta/llama-3.3-70b-instruct.toml:6 is attachment = true, sitting next
to text-only modalities, so the lab row is internally inconsistent rather than
already false. Lyceum rejects an image_url content part outright, so I kept the
override and the leading comment now explains why. Happy to drop it if the lab
row gets corrected to false instead.

bun validate passes.

[github-actions]

No actionable findings.