feat(nebul): add Nebul provider with 16 models and hourly sync module

[wynandhuizinga]

closes #6879
Adding Nebul as provider of models for opencode.

Nebul provider

Adds Nebul's Inference API — https://api.inference.nebul.io/v1 (OpenAI-compatible) — with 16 chat models and an hourly sync module driven by their public catalog endpoint.

Sources

Claim	Source
OpenAI-compatible base URL /v1, endpoints, NEBUL_API_KEY	https://docs.nebul.io/docs/inference-api/getting-started/quick-start
reasoning_effort (low/medium/high/max) is the only documented reasoning control; traces returned via reasoning_content (OpenAI-style, Nemotron, Qwen3 thinking), reasoning (GLM-5.x), or inline <thinking> (some Gemma)	https://docs.nebul.io/docs/inference-api/advanced-topics/reasoning
Per-model cost, cache_read, max_input_tokens, advertised reasoning_efforts (incl. Mistral-Medium-3.5's low|medium|high|max)	https://api.inference.nebul.io/model/info (live, unauthenticated)
models/alibaba/qwen3-30b-a3b-instruct-2507.toml dates, license, weights	https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507
models/mistral/ministral-3-14b-instruct-2512.toml (open weights, Apache 2.0, 256k usage example)	https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512

Reasoning controls policy

• Effort-only where the host advertises reasoning_efforts in /model/info (DeepSeek-V4.1-Flash low|high|max, GLM-5.3-Flash low|high|max, Mistral-Medium-3.5 low|medium|high|max — the advertised set supersedes the lab's none|high) or where the docs demonstrate it (gpt-oss-120b; GLM-5.3 uses the lab's always-on set low|high|max).
• No thinking toggles or budgets: Nebul's docs document no on/off or budget field, so toggle/budget options copied from lab or peer entries were dropped; reasoning-capable models without an advertised control use reasoning_options = [] with a leading source comment.
• The sync module preserves hand-authored non-effort options (toggle, budget_tokens) and replaces only the effort entry, matching the DigitalOcean pattern; covered by tests.

Sync deletion safety

• Existing TOMLs survive incomplete source data: the module prefers the authored base_model and falls back to existing cost/limit when /model/info transiently returns nulls, so hourly sync never deletes hand-authored files over temporary gaps (DeepInfra/Baseten pattern).
• Deletions follow only intentional removals: entries dropped server-side, deprecated/denylisted IDs, and non-chat surfaces; brand-new models must resolve a base model and price fully before entering the catalog.

Review fixes included

• models/mistral/ministral-3-14b-instruct-2512.toml: corrected open_weights to true, added license = "Apache 2.0" and the HF weights link (the same checkpoint is published on Hugging Face and marked open by NVIDIA/Infomaniak catalog entries); limit.output = 262_144 is sourced from the HF card's own usage example (max_tokens = 262144).

[github-actions]

Action items

• [high] [violation] providers/nebul/models/google/gemma-4-31B-it.toml:3 - Check: Every toggle needs a leading top-of-file wire-path comment. Why: This file declares type = "toggle" with no leading comment and no exact request field. Callers cannot tell how to turn reasoning on/off on Nebul’s OpenAI-compatible API, and hourly sync will not invent a header. Action: Add a top-of-file comment with Nebul’s exact toggle wire path (e.g. enable_thinking, thinking.type, reasoning.enabled) after verifying it, or drop toggle if this host has no on/off control.
• [high] [violation] providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml:3 - Check: Toggle wire comment + exact path; comments must be leading. Why: Toggle is present, but the only note is a mid-file “per alibaba” summary after base_model. That is not a leading wire comment, and sync strips non-header comments on rewrite. Action: Move a real # Toggle: … (and budget wire path if kept) above the first key for this file and the other toggle models (Qwen3.8-27B-FP8, Kimi-K3, all three Nemotron entries).
• [high] [violation] packages/core/src/sync/providers/nebul.ts:276 - Check: Sync must not destroy hand-authored non-effort reasoning controls. Why: buildReasoningOptions returns only [{ type: "effort", values }] whenever /model/info lists efforts, replacing the full authored set. That will drop toggle / budget_tokens on models such as Qwen3.8, Kimi-K3, and Qwen3.5 on the next sync that advertises efforts (DigitalOcean preserves non-effort options; Nebul does not). Action: Merge remote efforts with preserved non-effort options (and only replace the effort entry), matching the DigitalOcean/LLM Gateway pattern; add a test that toggle/budget survive when efforts are present.
• [high] [possible mistake] providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml:8 - Check: budget_tokens only when this host exposes a real reasoning-budget field. Why: Options are copied from Alibaba’s first-party entry, but Nebul is a multi-model OpenAI-compatible relay. Lab thinking_budget is not automatically available on every gateway; inventing budget_tokens without a Nebul wire field misstates caller controls. Action: Confirm Nebul accepts a reasoning budget on this route and document the field in a leading comment; otherwise remove budget_tokens and keep only controls this host actually exposes.
• [medium] [possible mistake] providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml:4 - Check: Relay reasoning baseline = lab + same-surface peers for this model. Why: DeepSeek’s own Flash entry and solid openai-compatible peers use toggle + effort low|high|max. Nebul authors effort only (from reasoning_efforts), so on/off may be missing if the host still forwards thinking.type / equivalent. Action: Verify whether Nebul can disable thinking on this model; if yes, add toggle plus a leading wire comment; if not, keep effort-only and note that the host has no separate off control.
• [medium] [possible mistake] models/mistral/ministral-3-14b-instruct-2512.toml:11 - Check: New lab metadata must reflect open-weight status accurately. Why: open_weights = false while the same checkpoint is published on Hugging Face and other catalog entries (e.g. NVIDIA) mark Ministral 3 14B Instruct as open weights. A wrong lab flag is inherited by every base_model host. Action: Confirm the HF release/license and set open_weights = true (and license if known) or cite why this instruct build is closed.
• [low] [possible mistake] .pr-review/pull-request.json:4 - Check: Data-changing PRs should cite provider pricing/docs/API mapped to claims. Why: PR body is empty while adding a provider, 16 priced models, two new lab entries, and a sync module, so costs, limits, and reasoning claims cannot be reviewed against first-party sources. Action: Add a PR body with direct Nebul /model/info or docs links (and lab sources for the new models/ files), stating what each citation supports.

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/nebul.ts:235 - Check: Sync modules must not skip source models when deleting existing local files for those IDs is unintentional (sync.md). Why: translateModel returns undefined whenever resolveBaseModel fails or pricing/context is null, and deleteMissing is left at the default (true). Unlike Baseten/Deep Infra peers, there is no existing.base_model fallback and no “update existing / skip create only” path. A temporary null cost field or an unresolved alias will drop the model from the desired set and hourly sync will delete the hand-authored TOML. Action: Preserve existing entries when pricing/context is incomplete (update only authoritative fields you have; keep cost/limit from existing). Prefer existing.base_model when present. Only return undefined for intentional catalog removals (ping/deprecated/non-chat), or set deleteMissing: false with a missingNotice if Nebul’s catalog is not authoritative for deletions.
• [medium] [possible mistake] providers/nebul/models/mistralai/Mistral-Medium-3.5-128B.toml:6 - Check: Relay reasoning_options must match the lab/same-surface baseline unless this host documents extra levels. Why: First-party providers/mistral/models/mistral-medium-2604.toml and solid peers (NVIDIA NIM, Scaleway, Evroc) use effort none|high. This file authors low|medium|high|max, which invents graded levels and drops the lab’s none off-switch. Action: Verify Nebul’s live reasoning_effort accepts and meaningfully applies low|medium|high|max on this model. If not, align to ["none", "high"] (or the exact set Nebul documents for Medium 3.5).
• [medium] [possible mistake] models/mistral/ministral-3-14b-instruct-2512.toml:22 - Check: New lab limit.output must be a real completion cap, not a copy of context. Why: output = 262_144 equals context. Peer provider caps for this model are much lower (e.g. NVIDIA 16_384, nano-gpt 32_768, OpenRouter 209_715, Infomaniak 25_600); the older Ministral lab sibling uses 8_192. An inflated lab default will inherit onto every base_model host. Action: Confirm the true max completion tokens from Mistral/HF docs and set limit.output (and any Nebul override) to that value.
• [low] [possible mistake] .pr-review/pull-request.json:4 - Check: Data-changing PRs should cite first-party pricing/docs/API evidence. Why: The body only says the provider is being added; costs, context overrides (e.g. DeepSeek/GLM 1M windows), and /model/info effort lists are material factual claims without mapped sources. Action: Add PR-body citations for Nebul pricing, /model/info, and the reasoning docs URL already used in TOML comments, stating what each URL supports.

[github-actions]

Action items

• [medium] [violation] providers/nebul/models/deepseek-ai/DeepSeek-V4.1-Flash.toml:3 - Check: Sync-owned TOML source/wire comments must be a leading header above the first key (AGENTS.md → Comments in TOML). Why: These files are rewritten by the Nebul hourly sync, which keeps only the leading comment block and drops everything after base_model. Effort/source notes on DeepSeek V4.1 Flash, Mistral Medium 3.5, GPT-OSS 120B, GLM-5.3, and GLM-5.3-Flash will disappear on the next sync. Action: Move those comments above base_model in deepseek-ai/DeepSeek-V4.1-Flash.toml, mistralai/Mistral-Medium-3.5-128B.toml, openai/gpt-oss-120b.toml, zai-org/GLM-5.3.toml, and zai-org/GLM-5.3-Flash.toml (same pattern as the Qwen/Nemotron leading headers).
• [low] [possible mistake] .pr-review/pull-request.json:4 - Check: Data-changing PRs should cite direct provider pricing/docs/API evidence in the PR body. Why: The body only says the provider is being added; it does not map costs, context overrides, or effort lists to /model/info, docs, or pricing pages, so material catalog values cannot be reviewed from the PR alone. Action: Add first-party citations in the PR body for Nebul pricing, /model/info effort lists, and any context overrides (and what each URL supports).

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/nebul.ts - Check: Provider-authored interleaved must describe this host’s reasoning side channel and survive sync. Why: The PR body documents traces via reasoning_content (OpenAI-style / Nemotron / Qwen3), reasoning (GLM-5.x), or inline tags, but none of the Nebul model TOMLs set interleaved, and translateModel never passes existing?.interleaved into factorBaseModel (unlike deepinfra/openrouter/chutes). Clients cannot discover the side channel, and any hand-authored interleaved will be dropped on the next hourly sync. Action: For each reasoner, author the correct Nebul wire form ([interleaved] field = "reasoning_content", or interleaved = true if the channel is only a generic/reasoning field the schema cannot name), and always preserve interleaved: existing?.interleaved (or an authored default) in the synced values object.
• [medium] [violation] packages/core/src/sync/providers/nebul.ts (buildReasoningOptions) - Check: Relay reasoning_options must match this host’s documented controls, not reintroduce lab/peer toggles. Why: After dropping undocumented toggles/budgets from the TOMLs, sync still keeps non-effort options whenever /model/info advertises efforts (preserved = existing…filter(type !== "effort") then append effort). That can rewrite files back to invalid Nebul shapes (e.g. toggle + effort) with no leading wire comment for a control Nebul does not document. Action: When reasoning_efforts is non-empty, replace reasoning_options with the advertised effort entry only; when empty, keep an explicit authored [] or other verified host-only set—do not merge lab-style toggle/budget back in.
• [medium] [possible mistake] models/mistral/ministral-3-14b-instruct-2512.toml - Check: Lab limit.output must be a real max-completion bound, not the context window. Why: This entry sets output = 262_144 from an HF usage example MAX_TOK = 262144, while other catalog hosts for the same checkpoint use much smaller completion caps (e.g. NVIDIA 16_384, OpenRouter/Kilo 209_715). Treating the full context size as limit.output overstates completion headroom for every base_model consumer. Action: Verify Ministral 3 14B’s actual max completion tokens from first-party or consistent peer evidence and set limit.output to that value (keep context = 262_144 if that remains correct).

[github-actions]

Action items

• [medium] [possible mistake] providers/nebul/models/moonshotai/Kimi-K3.toml:7 - Check: Relay reasoning_options must match this host’s real controls; [] only when there is affirmative no caller control (not uncertainty), using lab + same-surface peers as baseline. Why: Nebul is a multi-model relay whose docs center on reasoning_effort, and established peers of Kimi K3 (lab, Together, DeepInfra, Tinfoil, Berget) expose graded effort (low/high/max) on that kind of surface. Authoring [] here claims no control while peers and the host’s documented control disagree unless this model ID truly rejects effort. Action: Live-check whether reasoning_effort is accepted for moonshotai/Kimi-K3 on Nebul; if yes, set reasoning_options to the working effort list (typically low/high/max); keep [] only with evidence the parameter is rejected or ignored.
• [medium] [possible mistake] providers/nebul/models/meta-models/muse-glimmer-30b.toml:6 - Check: Same relay reasoning baseline / [] evidence bar. Why: Fireworks and Vercel peers document reasoning_effort as low/medium/high/xhigh for Muse Glimmer, and Nebul’s own docs treat reasoning_effort as the host control. Empty options contradict that baseline unless this route is proven effort-less. Action: Probe Nebul for this model; if effort works, author [{ type = "effort", values = ["low", "medium", "high", "xhigh"] }] (or the subset that actually changes behavior); retain [] only with rejection evidence.
• [medium] [possible mistake] providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml:6 - Check: Same relay reasoning baseline for models peers control via effort. Why: Same-surface peers (DeepInfra, OpenRouter) expose effort levels for Qwen3.8 27B (low/medium/xhigh, plus toggle where supported). Nebul drops both to [] because /model/info omits reasoning_efforts; catalog silence alone is weak if the shared OpenAI-compat reasoning_effort field still works. Action: Verify whether Nebul accepts reasoning_effort (and any on/off field) for this ID; author the intersection of what works (effort-only if no toggle), or keep [] only after confirming no control.

[github-actions]

Action items

• [high] [violation] providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml:5 - Check: Relay reasoning_options must not use [] from uncertainty when the host accepts caller controls. Why: This PR’s own later rationale says Nebul accepts reasoning_effort for reasoning-capable models (ignored only for non-reasoners), and patch 11 fixed Kimi-K3 / Muse Glimmer / Qwen3.8 for that reason — but Qwen3.5-397B, Gemma 4 31B IT, and all three Nemotron entries still ship reasoning_options = []. Lab/peers expose real controls (Qwen3.5: toggle/budget or effort; Gemma: toggle; Nemotron: toggle or effort with none). Empty here falsely claims no caller control on a multi-model OpenAI-compatible relay. Action: Author host-accurate effort sets for every remaining reasoner without advertised /model/info efforts (or prove per-model that Nebul ignores reasoning_effort for that ID), consistent with the three models already corrected.
• [high] [violation] packages/core/src/sync/providers/nebul.ts:276 - Check: Sync must not invent “no control” for new reasoners. Why: buildReasoningOptions returns undefined when reasoning_efforts is empty and there is no existing file; the shared runner then stamps reasoning_options = [] for any reasoning = true base (lab metadata never carries reasoning_options). Hourly sync will therefore auto-create new Nebul reasoners as empty-control entries — the AGENTS.md anti-pattern — instead of skipping them for hand authoring (Eden AI / Tinfoil pattern) or deriving a documented effort set. Action: For new chat reasoners with no advertised efforts and no authored options, skip translation (or fail closed for manual authoring); only write [] when the host affirmatively has no control for that model.
• [medium] [possible mistake] providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml:10 - Check: Effort values should match this host’s documented enum ∩ meaningful peer levels. Why: Comment cites Nebul’s documented reasoning_effort set and OpenRouter’s low|medium|xhigh, then authors only ["low", "medium"]. If Nebul’s public enum is low|medium|high|max, dropping high understates the control surface relative to the same docs used elsewhere in the PR. Action: Verify live/docs which levels Qwen3.8 actually honors on Nebul and include every accepted level (still excluding undocumented xhigh / lab toggle if unsupported).
• [low] [possible mistake] models/mistral/ministral-3-14b-instruct-2512.toml:18 - Check: limit.output must be a real max generation limit, not a sample max_tokens. Why: output = 262_144 is justified only by an HF usage example setting MAX_TOK = 262144 (same as context). That is often a request ceiling, not a verified generation limit; other hosts use much smaller output caps for this model. Action: Confirm Ministral’s actual max output tokens from Mistral/HF model card limits (not the sample script) and adjust if 262k is only a prompt/context figure.

[github-actions]

Action items

• [high] [violation] providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml:11 - Check: Relay reasoning_options must follow lab + same-surface peer controls for that model, not a host-wide invented effort enum. Why: Lab/Alibaba is toggle + budget_tokens; peers such as OpenRouter are toggle-only. /model/info does not advertise efforts for this model, and the file comment admits peers have “no effort values to intersect,” yet the entry authors effort = low|medium|high|max. That invents graded efforts the model/peers do not expose and misrepresents Nebul’s control surface. Action: Drop the invented enum. Use reasoning_options = [] unless you can show Nebul-specific evidence that graded reasoning_effort is meaningful for this model (or a real host toggle/budget wire path).
• [high] [violation] providers/nebul/models/google/gemma-4-31B-it.toml:12 - Check: Do not invent low/medium/high/max when lab/peers are toggle-only (or empty). Why: First-party Google and established peers (e.g. Crusoe, NVIDIA) use toggle only; Crusoe’s live probe even notes effort levels are inert and behave as on/off. Nebul does not advertise reasoning_efforts for Gemma, but the PR still pastes the host-wide enum. Action: Replace with reasoning_options = [] (or a documented toggle with a leading wire comment if Nebul actually exposes one). Do not publish graded efforts without model-specific host evidence.
• [high] [violation] providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml:11 - Check: Nemotron baseline is lab/peer toggle (not L/M/H/max). Why: NVIDIA first-party entries for Nano/Super are toggle-only; Omni is toggle (+ budget on NIM). These Nebul files invent low|medium|high|max while stating peers expose toggle only and /model/info lists no efforts. Action: For all three Nemotron entries (…Nano-30B…, …Super-120B…, …Nano-Omni…), remove the invented effort set and use [] unless Nebul documents/advertises real per-model effort values or a toggle wire path.
• [medium] [possible mistake] providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml:11 - Check: Effort values must be the intersection of peer/lab sets with this host’s real levels. Why: OpenRouter’s same-model set is low|medium|xhigh (plus toggle). Nebul drops toggle and xhigh, then adds high without /model/info advertising efforts for this ID. Action: Verify which reasoning_effort values Nebul actually accepts/effects for Qwen3.8-27B; author that set (likely low|medium and/or xhigh if supported), or [] if there is no caller control—do not invent high to fill a host-wide enum.

[wynandhuizinga]

Action items

• [high] [violation] providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml:11 - Check: Relay reasoning_options must follow lab + same-surface peer controls for that model, not a host-wide invented effort enum. Why: Lab/Alibaba is toggle + budget_tokens; peers such as OpenRouter are toggle-only. /model/info does not advertise efforts for this model, and the file comment admits peers have “no effort values to intersect,” yet the entry authors effort = low|medium|high|max. That invents graded efforts the model/peers do not expose and misrepresents Nebul’s control surface. Action: Drop the invented enum. Use reasoning_options = [] unless you can show Nebul-specific evidence that graded reasoning_effort is meaningful for this model (or a real host toggle/budget wire path).
• [high] [violation] providers/nebul/models/google/gemma-4-31B-it.toml:12 - Check: Do not invent low/medium/high/max when lab/peers are toggle-only (or empty). Why: First-party Google and established peers (e.g. Crusoe, NVIDIA) use toggle only; Crusoe’s live probe even notes effort levels are inert and behave as on/off. Nebul does not advertise reasoning_efforts for Gemma, but the PR still pastes the host-wide enum. Action: Replace with reasoning_options = [] (or a documented toggle with a leading wire comment if Nebul actually exposes one). Do not publish graded efforts without model-specific host evidence.
• [high] [violation] providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.toml:11 - Check: Nemotron baseline is lab/peer toggle (not L/M/H/max). Why: NVIDIA first-party entries for Nano/Super are toggle-only; Omni is toggle (+ budget on NIM). These Nebul files invent low|medium|high|max while stating peers expose toggle only and /model/info lists no efforts. Action: For all three Nemotron entries (…Nano-30B…, …Super-120B…, …Nano-Omni…), remove the invented effort set and use [] unless Nebul documents/advertises real per-model effort values or a toggle wire path.
• [medium] [possible mistake] providers/nebul/models/Qwen/Qwen3.8-27B-FP8.toml:11 - Check: Effort values must be the intersection of peer/lab sets with this host’s real levels. Why: OpenRouter’s same-model set is low|medium|xhigh (plus toggle). Nebul drops toggle and xhigh, then adds high without /model/info advertising efforts for this ID. Action: Verify which reasoning_effort values Nebul actually accepts/effects for Qwen3.8-27B; author that set (likely low|medium and/or xhigh if supported), or [] if there is no caller control—do not invent high to fill a host-wide enum.

1 — on the five remaining [] reasoners:
Fixed in d3bdd80 — all five now author effort sets. None of them has same-surface peer effort evidence to intersect (OpenRouter controls gemma-4-31b-it and nemotron-3-nano with a toggle, NVIDIA's first-party entry likewise, and qwen3.5-397b is toggle-only there too), so the tightest available evidence is Nebul's own docs, which state reasoning_effort is accepted on both endpoints and note the effort-to-budget mapping is model-specific guidance rather than a per-model contract. Each file now carries the documented host-wide enum low|medium|high|max with a leading comment saying exactly that. Since the docs also state non-reasoning models silently ignore reasoning_effort, the worst case on serving stacks that don't map a value is a no-op parameter.
2 — on sync auto-stamping [] for new reasoners:
Fixed in d3bdd80. translateModel now fails closed instead of emitting an empty options entry: when the resolved lab metadata says the model reasons (reasoning = true via the shared modelMetadata() accessor, the same pattern hyper/edenai use), the host advertises no reasoning_efforts, and no options are authored, it throws MissingReasoningOptionsError. The runner keeps the existing file, warns, and lists the model in the skipped notice (sync/index.ts:271-276, 459-462) — so new reasoners surface for hand-authoring, and existing files can never be overwritten with [] or deleted by the hourly job. A dry-run against the live catalog is clean (16 unchanged, no throws), and a test covers both the new-model and existing-file paths.
3 — on Qwen3.8-27B omitting high:
Fixed in d3bdd80 — now low|medium|high. high is in Nebul's documented value set and isn't contradicted by the peer evidence (OpenRouter advertises low|medium|xhigh); xhigh stays out because Nebul's docs never list it, and max stays out because no peer advertises it. The file comment spells this out.
4 — on Ministral's limit.output:
Keeping 262_144; d3bdd80 strengthens the citation. Mistral publishes no separate generation cap for this model — the first-party figures that exist are the card's 256k (= 262,144) context window, the vLLM section's default --max-model-len = 262144, and the card's own examples requesting max_tokens = 262144. The smaller figures on peer hosts (16_384 on NVIDIA, 25_600 on Infomaniak) are deployment/VRAM choices, not model limits. This also matches the catalog's convention for Mistral's current generation — mistral-large-2512, mistral-medium-2604 and devstral-2512 all set output = 262_144. Happy to switch if Mistral publishes an explicit max-output figure.

[github-actions]

Action items

• [high] [violation] providers/nebul/models/moonshotai/Kimi-K3.toml:1 - Check: Relay reasoning_options must match this host’s control surface consistently (AGENTS.md / audit skill: baseline = host + lab/peers; [] only for affirmative no control). Why: After the final “advertised reasoning_efforts is source of truth” policy, Qwen3.5/3.8, Gemma 4 31B, all three Nemotron entries, and GLM-5.3 correctly use reasoning_options = [] when /model/info lists no efforts. Kimi-K3 still authors low|high|max from Tinfoil despite the same empty advertisement, so the catalog claims caller effort control on one unadvertised reasoner and none on the others. Action: Align with the host policy: set reasoning_options = [] (and the same leading catalog comment as the other empty reasoners), or drop the advertised-only rule and author peer/lab-derived efforts for every unadvertised reasoner the same way—do not mix both.
• [high] [violation] providers/nebul/models/meta-models/muse-glimmer-30b.toml:1 - Check: Same host-wide reasoning-options consistency. Why: muse-glimmer still authors low|medium|high from Fireworks/Vercel while other models with empty reasoning_efforts are marked no-control. That misstates Nebul’s surface and will be preserved forever by sync when efforts stay empty. Action: Same fix as Kimi—either empty options with the catalog rationale, or apply peer-derived efforts uniformly to all unadvertised reasoners.
• [medium] [possible mistake] providers/nebul/models/zai-org/GLM-5.3.toml:1 - Check: Effort set vs lab sibling and host advertisement. Why: Lab/zhipuai peers use always-on low|high|max for both GLM-5.3 and GLM-5.3-Flash; Nebul keeps Flash at low|high|max but empties full GLM-5.3. That split is only valid if /model/info really returns efforts for Flash and an empty list for full GLM. Action: Re-check live reasoning_efforts for both IDs and either restore low|high|max on GLM-5.3 if Flash’s set is the real host surface, or leave [] only with a comment that cites the differing catalog rows for the two IDs.
• [low] [possible mistake] models/mistral/ministral-3-14b-instruct-2512.toml:1 - Check: Lab limit.output must be a real generation cap, not a context-sized max_tokens example. Why: output = 262_144 equals full context and is justified mainly by an HF usage example / peer flagships; providers that already host this model (e.g. Infomaniak) use a much smaller output bound. Overstated lab output will inherit to every base_model host. Action: Confirm Mistral’s actual max generation limit for Ministral 3 14B; if unknown, prefer a conservative documented cap (or omit a speculative full-context output) rather than copying context into limit.output.

[wynandhuizinga]

superseeded by #6961