feat(sync): add kenari sync adapter

[doedja]

Adds Kenari (kenari.id) to the hourly model sync workflow so its catalog stays current automatically instead of via hand-opened PRs.

How it works

• Fetches the public https://kenari.id/v1/models endpoint. No API key and no repo secret needed.
• Cost stays 0 by policy: Kenari bills in IDR from a prepaid wallet at its own catalog prices, so USD figures here would be a misleading currency conversion. The zero cost matches every existing hand-authored providers/kenari entry.
• Existing factored TOMLs are preserved exactly as authored. Limits keep inheriting the canonical base model.
• New models sync only when a canonical models/ metadata entry exists to inherit capability facts from (factorBaseModel). Kenari's dash IDs resolve to the dotted canon generically (grok-4-5 -> xai/grok-4.5), no per-model alias table.
• Models with no canonical entry are skipped and surfaced in the sync-report notice rather than authored from API data that has no name/release-date fields.
• Reasoning effort values come verbatim from the endpoint's reasoning_options array (per-model, e.g. "reasoning_options":["low","medium","high","xhigh","max"] on claude-fable-5). A model with "reasoning": true and no advertised efforts gets reasoning_options = [], matching the existing hand-authored entries.

Scope

This PR is the adapter only: packages/core/src/sync/index.ts, packages/core/src/sync/providers/kenari.ts, and its test. No TOMLs are added, changed, or deleted here. An earlier revision of this description carried a "first sync (included in this PR)" table, which was a dry-run result rather than a file list. Removed, since nothing in the diff writes a model file.

Where the catalog stands today

Measured 2026-08-10 against the live endpoint and this repo's default branch:

	count
Models the endpoint serves	50
providers/kenari/models/*.toml in this repo	38
Served but not in the repo	15
In the repo but no longer served	3 (grok-build-0-1, kimi-k2-6:free, kimi-k2-7-code:free)

The 15 is an upper bound on creations, not a prediction: any of them without a canonical models/ entry is skipped and reported instead. Most of that drift is also covered by #4393, so whichever of the two lands first shrinks the other's first run.

Reasoning effort values advertised by the endpoint

Re-checked against live on 2026-08-10, all still current:

Model	/v1/models reasoning_options
claude-fable-5, claude-sonnet-5	["low","medium","high","xhigh","max"]
gpt-5-6-luna, gpt-5-6-sol, gpt-5-6-terra	["none","low","medium","high","xhigh","max"]
grok-4-5	["low","medium","high"]
mimo-v2-5:free	none advertised (reasoning: true) -> reasoning_options = []

One thing worth disclosing about the effort values

Until 2026-08-10 the Kenari endpoint derived each model's reasoning_options from the openrouter slice of this repo's own feed, not from the model author. Had this adapter run before that date, it would have round-tripped models.dev data back into models.dev as though it were an independent observation, and any error in the OpenRouter slice would have been laundered into the providers/kenari rows as corroboration.

That is fixed at the source. The endpoint now reads the model author's own first-party slice when the author is a provider key this repo already carries, and falls back to the OpenRouter list otherwise. deepseek-v4-flash is the clearest case: it advertised ["high","xhigh"] (OpenRouter's list) and now advertises ["low","high","max"], which is what DeepSeek's own slice in this repo says. #4393 hand-corrects the rows that carried the old values.

So the adapter now reads a value the host derives from the author, and for a model whose author is not in this repo it still reads a value that traces back to the OpenRouter slice. That is the same provenance the hand-authored rows already had, and it is worth knowing rather than assuming every synced value is independent.

Verification

Current as of 2026-08-10:

• Both CI checks (review, validate) green on this branch
• The endpoint values in the table above re-measured against live

From when the adapter was written, against the catalog as it stood then:

• bun models:sync kenari twice: second run is a no-op (0 created, 0 updated, 0 deleted)
• bun validate passes
• bun run test: same 3 pre-existing failures as a clean checkout of dev, none related

[github-actions]

Action items

• [low] [violation] providers/kenari/models/claude-fable-5.toml:5 (also claude-sonnet-5.toml, gpt-5-6-luna.toml/sol/terra, grok-4-5.toml) - Check: The audit-reasoning-options skill requires explicit provider evidence for non-standard effort values (none, xhigh, max), and PR-body citations must state what each source proves. Why: The PR body cites the https://kenari.id/v1/models endpoint but never mentions reasoning_options at all, while the new files assert provider-specific effort sets like ["low","medium","high","xhigh","max"] (Claude 5), ["none","low","medium","high","xhigh","max"] (GPT-5.6), and ["low","medium","high"] (Grok 4.5). The adapter code sources these from the endpoint's reasoning_options field, but that mapping is not documented in the PR body, so the capability claims cannot be reviewed against the cited source without it. Action: Add a sentence to the PR body stating that the /v1/models response's reasoning_options array is the source for each new model's effort values (ideally noting the exact values the endpoint advertises per model), so the non-standard values are directly mapped to the cited endpoint.

[doedja]

Addressed the action item: the PR body now states that effort values are sourced verbatim from the endpoint's reasoning_options array, with a per-model table of the exact values https://kenari.id/v1/models advertises today (re-checked against the live endpoint before posting). Models advertising reasoning: true with no effort list get reasoning_options = [], same as the existing hand-authored entries.

[github-actions]

No actionable findings.

[doedja]

Opened #3380 as a data-only refresh of the kenari catalog, generated by running this adapter against the current dev branch. The live endpoint has drifted since this PR's snapshot (18 models added, 3 removed), which is the drift this adapter is meant to absorb hourly.

If #3380 lands first I will rebase this PR so its first sync becomes a no-op and it reduces to the adapter code plus the idempotence proof. Happy to also fit it into the labeling scheme from #3370 (model-sync, provider:kenari) if that is the direction sync PRs are heading.

[doedja]

Rebased onto current dev, and this is now pure code: 2 files, the adapter plus its registration, with zero data files in the branch.

I opened #4393 as the data-only refresh, generated by running this adapter on current dev. Once that lands, this adapter's first sync is a no-op, which was the thing I said I would do here rather than leaving you to review code and a data dump together.

The idempotence is measured, not assumed: running the adapter twice against the refreshed data gives 10 created, 0 updated, 3 removed on the first pass and 0 created, 0 updated, 0 deleted on the second, with validate.ts exiting 0.

One thing worth flagging rather than hiding: the live endpoint lists 50 models and the synced set is 45. The 5 without an entry have no canonical models/ base to inherit from, so the adapter skips them and reports them through skippedNotice. They are listed in #4393. Happy to hand-author those canonical entries separately if you want them covered.

Also still happy to take the model-sync and provider:kenari labels from #3370 whenever that scheme is ready.

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/kenari.ts:157 - Check: Relay reasoning_options must not invent “no control” from missing source data; preserve authored controls when the host omits them. Why: reasoningOptions() returns [] when reasoning === true and reasoning_options is omitted or empty, and the caller uses reasoningOptions(model) ?? existing.reasoning_options. Because [] is defined, the ?? fallback never runs, so any hand-authored efforts (e.g. DeepSeek high/xhigh, Gemini/Nemotron graded lists) are wiped whenever the endpoint leaves efforts unset. Peer adapters (OpenRouter, NanoGPT, Kilo) preserve existing options when the source omits controls; [] means affirmative no caller control, not uncertainty. Action: Return undefined when the API omits reasoning_options so existing entries are preserved; only write [] when the API explicitly returns an empty list (or for brand-new models with no prior entry via the runner’s preserve path).
• [medium] [possible mistake] packages/core/src/sync/providers/kenari.ts:148 - Check: New Kenari models must resolve base_model reliably to canonical models/ metadata. Why: resolveKenariBaseModel only tries resolveCanonicalBaseModel(owned_by + id|dotted id). That requires owned_by to be a CANONICAL_PROVIDER_PREFIXES key (no zhipuai, bare-lab typos, etc.) and never falls back to bare unique filename resolution. Requesty/NanoGPT/EmpirioLabs use resolveModelMetadataBaseModel or ID-derived candidates so bare aggregator IDs still factor. Mis-tagged or unknown owned_by silently skips creatable models even when models/<lab>/<id>.toml exists. Action: After the owned_by attempts, fall back to resolveModelMetadataBaseModel on the raw and dash-to-dot IDs (and add any Kenari owned_by aliases you observe, e.g. zhipuai → zai/zhipuai).
• [medium] [violation] packages/core/src/sync/providers/kenari.ts:165 - Check: Sync code must locate models/ independently of process cwd (see other providers’ MODELS_DIR). Why: canonicalName() reads path.join("models", …) from the working directory. Other adapters resolve metadata via import.meta.dirname. If sync is not launched from the repo root, :free display names fail open (freeName returns undefined) and new free variants lose the "(Free)" name override. Action: Resolve the metadata root from import.meta.dirname (same pattern as openrouter.ts / requesty.ts) and read canonical names from that absolute base.
• [medium] [possible mistake] packages/core/src/sync/providers/kenari.ts:1 - Check: New sync adapters with custom ID normalization and reasoning mapping need focused regression tests (sync.md / existing sync.test.ts coverage for EmpirioLabs, LLM Gateway, Hyper, etc.). Why: Dash↔dot resolution, :free handling, existing-vs-new factoring, and the omitted-vs-empty reasoning behavior are easy to regress and are untested in this PR. Action: Add unit tests for resolveKenariBaseModel / buildKenariModel covering dotted canon IDs, :free naming, preservation of authored reasoning_options when the API omits efforts, and zero-cost factored output.

[doedja]

All four review items are fixed, each with a test that fails when the fix is reverted.

• reasoningOptions() returns undefined when the endpoint omits efforts, so both ?? existing sites and the runner's preserveReasoningOptions work. Only an explicitly empty list writes [].
• resolveKenariBaseModel falls back to resolveModelMetadataBaseModel on the bare and dotted IDs when owned_by is not a canonical prefix. It still skips when the match is ambiguous.
• canonicalName reads from a MODELS_DIR derived from import.meta.dirname, matching openrouter and requesty, so :free names no longer depend on cwd.
• reasoning_options parses as strings and filters to the catalog effort set, deduped, so one unknown value no longer throws and kills the whole sync. A list with no known values is treated as silence rather than as "no control".

Also dropped modalities and tool_call from the schema, which were declared but never read.

9 tests added. Full sync.test.ts is 122 pass, 1 fail, and that failure (DeepInfra preserves live modalities for new base models) also fails on clean dev. tsc --noEmit reports 36 errors on this branch and 36 on clean dev, none in kenari files.

The base-resolution fix changed the data, so I regenerated #4393: it now covers 48 of the 50 live models instead of 45. ling-3-0-flash:free and north-mini-code:free still skip, since neither has canonical metadata to inherit from.

Happy to take the model-sync and provider:kenari labels from #3370 whenever that scheme is ready.

[github-actions]

Action items

• [high] [possible mistake] packages/core/src/sync/providers/kenari.ts:148 - Check: New models must resolve a canonical models/ entry reliably (same-surface peers use bare-ID fallbacks). Why: resolveKenariBaseModel only tries resolveCanonicalBaseModel("${owned_by}/${id}") (plus dash→dot). That requires Kenari’s owned_by to be an OpenRouter-style lab prefix (openai, x-ai, z-ai, …). Values like system, kenari, or zhipuai never match, so every create is skipped. Idempotence on existing TOMLs only exercises the existing.base_model update path, and there is no test for the create/resolve path—so a broken owned_by mapping can ship unnoticed. Action: After the prefixed attempts, fall back to resolveModelMetadataBaseModel on the bare/dotted id (as cortecs/hyper/requesty do); add unit tests for representative Kenari ids/owned_by pairs; confirm against live /v1/models that new rows actually resolve.
• [medium] [violation] packages/core/src/sync/providers/kenari.ts:104 - Check: Sync must stay resilient when the provider API adds fields outside the catalog enum. Why: reasoning_options: z.array(EffortValue) makes parseModels throw on any unknown effort string, failing the entire Kenari hourly job instead of syncing the rest of the catalog. Peer adapters (crossmodel, hyper, empiriolabs) accept strings and filter to known efforts. Action: Parse efforts as z.array(z.string()) (or similar) and keep only values in the catalog effort set when building reasoning_options.
• [low] [possible mistake] packages/core/src/sync/providers/kenari.ts:165 - Check: Provider modules must resolve models/ via a stable path (not process cwd). Why: canonicalName reads path.join("models", …), unlike every other adapter’s import.meta.dirname → repo models/ root. Free-tier display names ("${base} (Free)") silently drop if cwd is not the repo root. Action: Use the shared MODELS_DIR / modelMetadata() pattern and cover freeName in tests.

[github-actions]

Checking one more interaction: whether the create path and preserveReasoningOptions force [], and whether the factored path drops provider/experimental.

[github-actions]

Checking how resolved existing models expose base_model and whether the adapter mishandles it.