Dev

[wynandhuizinga]

Add https://nebul.com as a provider with its in-scope chat models (12 today), plus an hourly sync module.

Supersedes #6876 (rebased cleanly onto current dev; resolves its remaining review items). The Ministral 3 14B lab file landed upstream in the meantime, so the only new lab entry is Alibaba's Qwen3 30B A3B Instruct.

Closes #6879.

Sources

Data	Verified against
Reasoning effort values	live GET https://api.inference.nebul.io/model/info "reasoning_efforts" is the per-model source of truth (GLM-5.3-Flash: low|high|max; GLM-5.3: empty list; Mistral Medium 3.5: low|medium|high|max)
gpt-oss-120b high/medium/low via reasoning_effort	Nebul docs (only model the docs demonstrate)
Wire fields	Nebul docs: reasoning_effort param, no toggle, no budget; traces in message.reasoning
Lifecycle	model_info.superseded_by_model_name in /model/info (GLM-5.1/5.2-FP8 → GLM-5.3; Qwen3-VL-235B-A22B-Thinking → Qwen3.5-397B-A17B)

Catalog scope

• General chat models only: embeddings and rerankers (by mode/model_type), specialized document-OCR models (name-scoped), entries the host flags via display_tags (Guard Model / Content Safety / Private / Internal), and server-side-superseded entries all skip silently. No name lists beyond the OCR scope rule exist in the module.
• Private/Internal-tagged entries (e.g. the approval-gated Google and Ministral IDs dropped during review) are excluded durably by this tag filter, independent of server-side timing. Two early-access models dropped during review carry no tag, so their durable exclusion still depends on server-side removal.

Reasoning controls policy

• Effort-only where the host advertises values in /model/info or the docs demonstrate them (the models listed above). All other reasoners use reasoning_options = [] with a sourced comment; no toggles, no budgets.
• Sync preserves authored [] exactly; if nothing is advertised and nothing authored, new syncs fail closed with a missing-reasoning-controls error instead of inventing values.

Sync deletion safety

• Deletes only happen when the host still serves enough entries (deleteDropped ≥10 live models / ≥50% retention), preserving existing catalog models otherwise.
• Whole-catalog faults (empty response, or no entry matching the chat-model filter) fail closed in parseModels before any file is written or deleted.

[github-actions]

Action items

• [high] [possible mistake] providers/nebul/models/moonshotai/Kimi-K3.toml:4 - Check: Provider reasoning / reasoning_options / interleaved must match this host’s actual reasoning surface. Why: The leading comment states Nebul marks this ID with supports_reasoning = false, but the entry still inherits reasoning = true from moonshotai/kimi-k3, sets reasoning_options = [], and declares [interleaved] field = "reasoning_content". That combination claims an always-on reasoner with a reasoning side channel while the same file asserts the host does not support reasoning. Action: Verify the live /model/info flags and chat behavior for moonshotai/Kimi-K3. If the host truly does not reason, set reasoning = false and remove reasoning_options and interleaved. If it does reason (always-on, no controls), drop the supports_reasoning = false claim and keep [] only with accurate evidence.
• [medium] [violation] packages/core/src/sync/providers/nebul.ts:144 - Check: Skipped remote IDs and skippedNotice must describe the real skip reason. Why: sourceID returns every in-scope chat model that translateModel skips, including brand-new entries missing price or max_input_tokens, not only unresolved base_model targets. skippedNotice always blames missing lab metadata (need hand-authored base_model targets), so pricing/limit gaps are misreported and any future missing-model issue wiring would open the wrong class of issues. Action: Split skip reasons (unresolvable base vs incomplete serving data) in sourceID / notices, and only route true missing-lab cases through the base_model hand-author path (optionally via missingModelID like other multi-lab relays).

[github-actions]

Action items

• [high] [violation] providers/nebul/models/moonshotai/Kimi-K3.toml:11 - Check: reasoning_options = [] means the model reasons with no caller control; do not mark a non-reasoner that way, and do not attach an interleaved reasoning side-channel unless traces are actually returned. Why: The leading comment states the live catalog marks this ID supports_reasoning = false, but the file still inherits reasoning = true from moonshotai/kimi-k3, sets reasoning_options = [], and declares [interleaved] field = "reasoning_content". Those three claims cannot all be true: either Nebul does not reason here (override reasoning = false and drop reasoning_options + interleaved) or it does reason (remove the supports_reasoning = false claim and keep only controls/trace fields that match the host). Action: Reconcile against Nebul /model/info + a live completion for this ID; author the matching override set and fix the header comment so it no longer contradicts the TOML.