Add Nebul as provider (4 curated chat models + sync)

[wynandhuizinga]

Add https://nebul.com as a provider with four curated chat models, plus an hourly sync module.

Supersedes #7400 (same provider, deliberately trimmed to the four flagship models below for a smaller review surface). Ref #6879.

Models added (all base_model + override-only against existing lab metadata — no new lab files needed):

Served ID	Lab base	Cost (in/out/cache-read, USD/MTok)	Context
zai-org/GLM-5.3	zhipuai/glm-5.3	1.47 / 4.62 / 0.35	1,048,576
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16	nvidia/nemotron-3-super-120b-a12b	0.32 / 0.69 / 0.08	1,000,000
Qwen/Qwen3.5-397B-A17B	alibaba/qwen3.5-397b-a17b	0.63 / 3.78 / 0.15	262,144 (inherited)
mistralai/Mistral-Large-3-675B-Instruct-2512	mistral/mistral-large-2512	0.60 / 1.73 / 0.15	inherited

Curation policy

The sync treats the catalog as authoritative for these four entries' live cost/context, but it can never grow or shrink the shipped set:

• skipCreates: true — only the four hand-authored models ship; any other in-scope chat model is reported via the skipped notice and left out.
• deleteMissing: false — a curated model absent from the catalog (e.g. zai-org/GLM-5.3 has been intermittently absent) is retained and surfaced via missingNotice rather than silently deleted.
• trackMissingModels: false — no missing-model issues for in-scope chat models this provider deliberately does not curate.

Sources

Data	Verified against
Costs and per-model context	Live GET https://api.inference.nebul.io/v1/model/info (input_cost_per_1m_tokens, output_cost_per_1m_tokens, cache_read_input_cost_per_1m_tokens, max_input_tokens), re-checked 2026-09-24
Nemotron 3 Super & Qwen3.5 397B effort values	Live probes (POST /v1/chat/completions, 2026-09-23): reasoning_effort accepts none|minimal|low|medium|high|xhigh|max — all levels return 200; thinking is on by default, none disables it. The catalog's supports_reasoning = false for both IDs is a false negative
GLM-5.3 effort values low|high|max	Live catalog reasoning_efforts + parity with the lab (providers/zai) and OpenRouter entries for the same base model
Wire fields	Nebul docs: reasoning_effort param only — no on/off toggle, no reasoning budget; traces arrive in reasoning_content (hence interleaved)

Reasoning controls policy

• Effort-only, per the probe/doc evidence above; no toggles, no budgets invented for this API.
• Sync preserves authored, probe-verified options exactly (they outrank the advertised reasoning_efforts, which have been observed wrong on this host). If a reasoner advertises nothing and has nothing authored, sync fails closed with a missing-reasoning-controls error instead of inventing values.

Sync safety

• All four files carry leading header comments and are re-verified by sync rather than blindly rewritten: sync keeps existing entries when the host's pricing/context is temporarily null and when a served alias stops resolving to lab metadata.
• Whole-catalog faults fail closed in parseModels before any file is written or deleted: empty response, no entry matching the chat-model filter, or a catalog smaller than half the known-good in-scope size.
• Out-of-scope entries (embeddings/rerankers by model_type, document-OCR by name, display_tags Guard Model / Content Safety / Private / Internal, server-side-superseded IDs) skip silently.

Checklist

• Compliant logo.svg (currentColor, 0 0 24 24 viewBox)
• Non-lab models use base_model; all four lab files already exist upstream
• Provider files override-only (USD/MTok costs, reasoning options, limits)
• bun validate passes
• bun test packages/core/test/nebul.test.ts — 21/21 pass
• bun packages/core/script/sync-models.ts nebul --dry-run — 4 unchanged, 0 created/updated/removed

[github-actions]

Action items

• [critical] [violation] packages/core/src/sync/providers/nebul.ts:9 - Check: Sync fetch URL must match the authoritative catalog endpoint. Why: Code fetches https://api.inference.nebul.io/model/info but PR body, TOML headers, and provider.toml api all use https://api.inference.nebul.io/v1/model/info; the missing /v1 will fail the hourly sync. Action: Fix API_ENDPOINT to the /v1/model/info URL and verify with a dry-run.
• [high] [possible mistake] providers/nebul/models/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml:12 - Check: Relay effort sets must copy lab + same-surface peers; extra levels need host docs or live meaningful effect. Why: Lab is toggle-only and peers are narrower (openrouter toggle+low|medium+budget, requesty none|low|medium|high|max+budget, edenai none|low|medium, kenari low|medium); no peer shows minimal+xhigh for this model, and Qwen/Qwen3.5-397B-A17B.toml:12 similarly exceeds lab (toggle+budget), openrouter (toggle only), and ovhcloud/digitalocean (none|low|medium|high) on minimal/xhigh/max. A 200 OK acceptance probe alone does not prove graded effect or that none disables traces. Action: Provide docs excerpt listing supported reasoning_effort values and probe excerpts showing differing traces / none disabling, or narrow to the verified set.
• [high] [violation] packages/core/src/sync/providers/nebul.ts:191 - Check: Sync must fail closed on unverified reasoning controls, not auto-create from advertised lists. Why: buildReasoningOptions falls back to advertised reasoning_efforts for new entries, but the PR proves the catalog unreliable (supports_reasoning=false false negatives, Mistral-Medium-3.5 advertised low|medium|high|max but only high works); with 4 of ~12 catalog models committed, hourly sync will auto-create ~8 unreviewed models with potentially wrong controls. Action: Throw MissingReasoningOptionsError for new reasoners without authored probe-verified options (or set skipCreates/missingModelID flow), or commit all in-scope models with verified options.

[wynandhuizinga]

All three findings addressed in ddf28e82.

1. Endpoint (nebul.ts:9): live-verified 2026-09-23: both https://api.inference.nebul.io/model/info and https://api.inference.nebul.io/v1/model/info return 200 with the identical catalog (with and without auth), so the sync would not have failed — but the bare path was inconsistent with provider.toml and the PR body. Aligned to /v1/model/info.

2. Effort sets: narrowed both nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml and Qwen/Qwen3.5-397B-A17B.toml to none|low|medium|high|max, matching the widest same-surface peer baseline (requesty). minimal/xhigh will return in a follow-up with per-level distinct-effect probe evidence.

3. Fail-open advertised-efforts fallback (nebul.ts:191): fixed. New reasoners without hand-authored reasoning_options now throw MissingReasoningOptionsError (runner warns + skips); the catalog's advertised list is never copied into a synced entry. Dry-run against the live catalog after the change: 0 created, 0 updated, 0 removed, 4 unchanged, with 9 reasoners skipped pending hand-authored probe-verified options (incl. mistralai/Mistral-Medium-3.5-128B, whose advertised low|medium|high|max is disproven by the recorded probes — only high works).

bun test (nebul suite 20/20, incl. a new fail-closed test) and validate both pass.

[github-actions]

Action items

• [high] [violation] providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml:13 - Check: Reasoning-options baseline must be lab + same-surface peer set; extra levels beyond lab/peers need host docs or live meaningful effect. Why: Lab (providers/alibaba/models/qwen3.5-397b-a17b.toml) is toggle+budget_tokens and established relay peers (providers/ovhcloud, providers/scaleway, providers/crof, providers/digitalocean/models/qwen3.5-397b-a17b.toml) are none|low|medium|high — none include max. A 200 OK probe alone does not prove a distinct reasoning effect and contradicts the file's own narrowing rationale for minimal/xhigh. Action: Drop max to ["none", "low", "medium", "high"] to match peers, or keep it only with distinct-effect evidence (e.g., differing reasoning-token counts/outputs for max vs high plus doc reference) cited in the PR body/header.

[wynandhuizinga]

Qwen3.5 finding addressed in cde5df80: dropped max — providers/nebul/models/Qwen/Qwen3.5-397B-A17B.toml is now ["none", "low", "medium", "high"], matching the established relay-peer baseline (ovhcloud/scaleway/crof/digitalocean all carry none|low|medium|high; the lab entry itself is toggle+budget_tokens). Header comment updated to cite the peer set and note that minimal/xhigh/max lacked distinct-effect evidence. (NVIDIA-Nemotron-3-Super-120B-A12B-BF16.toml keeps max deliberately: requesty's entry for the same model carries none|low|medium|high|max.) bun test (nebul suite 20/20) and validate pass.

[github-actions]

No actionable findings.

[github-actions]

Action items

• [medium] [violation] packages/core/src/sync/providers/nebul.ts:137 - Check: Sync parsing must not fail the whole run on an unused source field. Why: translateModel deliberately never copies reasoning_efforts because probes proved the catalog list unreliable, yet EffortValues strict enum makes NebulResponse.parse throw on any future unknown value, failing hourly sync and blocking cost/context refreshes for the curated models. Action: Loosen the field to accept unknown strings (e.g. z.array(z.string())) since it is never synced, and update the rejects unknown reasoning effort values test in packages/core/test/nebul.test.ts to assert unknown values are accepted/ignored.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.