chore(kenari): refresh model catalog from the live endpoint

[doedja]

Data-only refresh of the Kenari catalog against the live public
https://kenari.id/v1/models endpoint. 16 models added, 8 removed, 6
already-merged rows corrected. No adapter or sync code in this branch.

Related: #5727 (Kenari sync adapter) is still open from 2026-08-28. This
refresh is against current dev so the picker catalog can catch up
without waiting on that adapter.

Added (16)

Every field below comes from GET https://kenari.id/v1/models, which is
public and needs no key.

model id	base_model	reasoning_options from the endpoint
deepseek-v4-1-flash	deepseek/deepseek-v4.1-flash	toggle + ["low","high","max"]
gemini-3-8-flash	google/gemini-3.8-flash	["low","medium","high"]
gpt-5-4	openai/gpt-5.4	["none","low","medium","high","xhigh"]
gpt-6-astra	openai/gpt-6-astra	["low","medium","high","xhigh","max"]
hy4-preview	tencent/hy4-preview	["none","low","high"]
laguna-s-2-1:free	poolside/laguna-s-2.1	toggle (Poolside first-party)
laguna-xs-2-1:free	poolside/laguna-xs-2.1	toggle (Poolside first-party)
muse-spark-1-2	meta/muse-spark-1.2	["minimal","low","medium","high","xhigh"]
muse-spark-1-2-contributor	meta/muse-spark-1.2	same (name override, same as Meta first-party)
muse-spark-1-2-contributor:free	meta/muse-spark-1.2	same
muse-spark-1-3	meta/muse-spark-1.3	["minimal","low","medium","high","xhigh","max"]
muse-spark-1-3-contributor	meta/muse-spark-1.3	["minimal","low","medium","high","xhigh"]
muse-spark-1-3-contributor:free	meta/muse-spark-1.3	same
nemotron-3-ultra-550b-a55b:free	nvidia/nemotron-3-ultra-550b-a55b	toggle + ["medium","high"]
north-mini-code:free	cohere/north-mini-code-1-0	["none","high"] (Cohere first-party)
qwen3-8-flash	alibaba/qwen3.8-flash	toggle + ["low","medium","xhigh"]

Contributor rows follow Meta's own files: same base_model, display
name override. north-mini-code:free folds onto
cohere/north-mini-code-1-0.

Corrected (6 existing rows)

model id	was	now	why
deepseek-v4-flash	["high","xhigh"]	toggle + ["low","high","max"]	DeepSeek lab + Kenari $.reasoning.enabled
deepseek-v4-pro	["high","xhigh"]	toggle + ["high","max"]	same
nemotron-3-super-120b-a12b	effort ["low","medium"]	toggle + ["low","medium"]	NVIDIA toggle; Kenari forwards $.reasoning.enabled
nemotron-3-super-120b-a12b:free	effort ["low","medium"]	same	same
nemotron-3-ultra-550b-a55b	effort ["medium","high"]	toggle + ["medium","high"]	same
gpt-oss-20b	full inline TOML	base_model = "openai/gpt-oss-20b"	lab entry exists now

Removed (8)

Confirmed absent from the live response:

deepseek-v4-flash:free, gemini-3-5-flash, glm-5-1, grok-build-0-1,
kimi-k2-6:free, kimi-k2-7-code:free, mistral-large:free,
nemotron-3-nano-30b-a3b

Skipped

Video endpoints, left out of this picker list:

gemini-omni-flash, veo-3.1-lite

No canonical models/ entry to inherit from, so not authored from the
feed (the endpoint still has no name or release-date data):

agnes-2-0-flash:free, agnes-2-5-flash:free, grok-imagine-image,
grok-imagine-image-quality, kokoro-tts, mimo-v2-5-tts,
minimax-speech-2-8-hd, minimax-speech-2-8-turbo, nano-banana-2,
nano-banana-2-lite, nano-banana-pro, nex-n2-5-pro:free

grok-imagine-image is not folded onto grok-imagine-image-2.0. That
would give two live IDs the same lab entry.

Image (gpt-image-2, grok-imagine-image-2-0) and audio
(whisper-large-v3-turbo, gemini-3-1-flash-tts) stay.

Kenari is a multi-model OpenAI-compatible relay. It forwards
$.reasoning.enabled and reasoning_effort. It does not expose Alibaba
thinking_budget, so budget_tokens is not authored.

Cost stays 0

Kenari bills a prepaid wallet in Indonesian rupiah. A USD per-token
figure here would be a currency conversion that drifts. Same treatment
as the previous merged refreshes.

Verification

• Source: GET https://kenari.id/v1/models on 2026-09-14 (81 live models)
• bun ./packages/core/script/validate.ts exits 0 (67 Kenari entries after merge)

[github-actions]

Action items

• [high] [violation] providers/kenari/models/nemotron-3-ultra-550b-a55b.toml:2 - Check: Relay reasoning_options must not be [] from missing endpoint fields when lab/peers expose controls. Why: This PR replaces prior Kenari effort ["medium", "high"] (aligned with OpenRouter peers) with [] because the catalog omitted options. Lab NVIDIA uses toggle; OpenRouter uses toggle + medium/high. Empty means no caller control, not “not published.” Same regression on the free twin. Action: Restore real controls (at least prior/openrouter effort, and toggle if Kenari forwards on/off); only keep [] with affirmative proof this host has no control. Apply the same fix to nemotron-3-ultra-550b-a55b:free.toml, nemotron-3-super-120b-a12b.toml, and nemotron-3-super-120b-a12b:free.toml.
• [high] [violation] providers/kenari/models/qwen3-8-flash.toml:2 - Check: Multi-model relay of a controlled reasoner must copy lab/same-surface peer options, not invent [] from catalog gaps. Why: base_model = "alibaba/qwen3.8-flash" is a reasoner; Alibaba first-party and peers (OpenRouter, Vercel) expose toggle and often budget_tokens / effort. Authoring reasoning_options = [] because Kenari’s models list published none is the documented anti-pattern and hides caller controls. Action: Set options from Alibaba + openai-compatible peers that match what Kenari actually forwards (typically toggle + budget_tokens and/or lab effort); do not leave [] without host proof of no control.
• [high] [violation] providers/kenari/models/north-mini-code:free.toml:3 - Check: Relay baseline = lab + peers for that model. Why: PR rationale is reasoning: true, none published → []. Cohere first-party is effort ["none", "high"]; other free hosts keep controls. [] here is uncertainty, not affirmative no control. Action: Use [{ type = "effort", values = ["none", "high"] }] (or Kenari-verified equivalent); only use [] if the host truly has no reasoning parameter.
• [high] [violation] providers/kenari/models/laguna-s-2-1:free.toml:3 - Check: Same relay [] rule for models with lab toggle. Why: Poolside first-party Laguna S/XS 2.1 uses toggle (chat_template_kwargs.enable_thinking). Catalog omission is not evidence Kenari has no control. Action: Match lab/peers (toggle with a leading wire comment if supported, or verified effort); apply the same to laguna-xs-2-1:free.toml. Do not use [] solely because the endpoint omitted options.

[doedja]

Addressed the reviewer action items on the eight [] files.

Kenari is a multi-model relay. It forwards $.reasoning.enabled and reasoning_effort. Empty lists from a missing /v1/models field are gone.

• Nemotron Super/Ultra (paid and free): toggle plus the prior OpenRouter effort lists (low/medium on Super, medium/high on Ultra). NVIDIA first-party is toggle-only; Kenari also forwards effort, so both stay.
• qwen3-8-flash: toggle plus Alibaba lab effort low/medium/xhigh. No budget_tokens; Kenari has no thinking_budget field.
• north-mini-code:free: Cohere first-party effort none/high (no toggle; none is the off switch).
• Laguna S/XS 2.1 free: Poolside first-party toggle.

bun ./packages/core/script/validate.ts still exits 0.

[github-actions]

Action items

• [high] [violation] providers/kenari/models/deepseek-v4-1-flash.toml:3 - Check: DeepSeek V4 on a relay that exposes on/off must copy lab toggle + effort (not effort alone). Why: Lab Flash/V4.1 Flash is toggle + low/high/max; this PR states Kenari forwards $.reasoning.enabled and adds that toggle for Nemotron/Laguna/Qwen, but the three DeepSeek rows only author effort. Callers lose the documented on/off control. Action: Add [[reasoning_options]] type = "toggle" (with a leading # Toggle: $.reasoning.enabled = true|false comment) alongside the existing effort list on deepseek-v4-1-flash.toml, deepseek-v4-flash.toml, and deepseek-v4-pro.toml (high/max for Pro).

[doedja]

Latest reviewer item: DeepSeek V4 on this relay needs lab toggle plus effort, not effort alone.

deepseek-v4-1-flash, deepseek-v4-flash, and deepseek-v4-pro now author toggle ($.reasoning.enabled) next to the existing effort lists (low/high/max on Flash, high/max on Pro). Matches DeepSeek first-party.

bun ./packages/core/script/validate.ts still exits 0.

[github-actions]

No actionable findings.