omp-kenari

Kenari ID provider for the omp coding agent.

Export KENARI_API_KEY and the plugin imports Kenari's live model catalogue, routes the judge model role to Kenari's Jev System One endpoint, and adds tools for the capabilities a provider registration cannot reach. There is no config file to write.

Install

omp plugin install omp-kenari        # from npm
omp plugin link /path/to/omp-kenari  # local checkout
export KENARI_API_KEY=...

The host executes src/*.ts directly — there is no build step, so the published package is exactly the tree you develop against.

What it registers

Provider	API	Contents
kenari	openai-completions	Every chat model from GET /v1/models, discovered live
kenari-judge	typesafe	jev-1-13-free, served at POST /v1/systemone

Both resolve the credential through the KENARI_API_KEY environment variable, so the key is never copied into omp's config or credential store.

The judge role is pointed at kenari-judge/jev-1-13-free only when the role is unconfigured. An explicitly configured judge is left alone, and the role is not touched without a key — the override is runtime-only and never persisted.

Tools

Tool	Endpoint	Purpose
kenari_search	POST /v1/web/search	Live web search; ranked titles, URLs and snippets
kenari_fetch	POST /v1/web/fetch	One URL as clean Markdown, plus title and outbound links
kenari_embed	POST /v1/embeddings	Embed text (bge-m3 by default)
kenari_rerank	POST /v1/rerank	Rerank candidates by relevance (bge-reranker-base)

Search, fetch and rerank are billed to the account's balance; embeddings are effectively free at this scale. All four resolve the credential through omp's registry, so a key in config.yml works the same as one in the environment.

Commands

/kenari   # plan, quota windows, reset times, model counts, FX rate, free lane

A one-line version of the same summary is shown when a session starts. Both read GET /v1/account/quota and /api/public/pricing; neither is required for the providers to work.

The free-lane figures are tier-dependent and the tier is chosen by account state, not printed from the first entry: a subscriber sees the plan tier (15 req/min), not the new-account tier (5 req/min). A window Kenari omits is unlimited, not missing — zero means unlimited in their API.

Settings

omp plugin config list omp-kenari
omp plugin config set omp-kenari filterPayPerUse false

Key	Default	Effect
discoveryEnabled	true	Import the live catalogue. When off, only the judge provider is registered.
filterPayPerUse	true	Keep only models that are free or included in your account's own plan.
autoJudge	true	Register the judge provider and default the judge role to it.
toolsEnabled	true	Register the four Kenari tools.
startupNotice	true	Show the account summary when a session starts.

Plan membership is per-plan, not per-catalogue. Every paid plan includes claude-sonnet-5, but claude-opus-4-7 is in no plan at all and gpt-5-6-sol only in the top two tiers. The plugin reads the account's actual plan from GET /v1/account/quota and filters against that plan's model list, so models the account cannot use without paying per token are hidden — 25 of them on a Studio plan, leaving 47.

When the plan is unknown (no subscription, or the quota endpoint failed), plan membership is unknown and every model is listed: hiding a model the account may already pay for is worse than showing one that costs per token.

Catalogue-shaping settings are applied when the model list is fetched, and omp caches that list per provider for 24 hours. After changing a setting, drop the cache so the new shape is fetched:

omp models refresh   # or delete ~/.omp/agent/models.db*

Notes

• Only chat models are registered. Kenari also serves image (/v1/images/generations), video, speech, transcription and embedding ids, but omp offers no way for an extension to register a non-chat model kind, so they are intentionally left out rather than registered as chat models they are not. The embedding and rerank ids remain reachable through kenari_embed and kenari_rerank.
• Prices arrive as micro_idr_per_1m_tokens (whole rupiah × 1e6). omp's cost model is USD-only, so they are converted with the live usd_idr_rate from /api/public/pricing (falling back to a documented constant) and used only for the relative-cost readouts omp derives. They are not a billing source of truth.
• reasoning_options maps onto omp's effort levels one-to-one, minus none: omp expresses "do not think" by omitting reasoning_effort, so none is not offered as an effort level. A reasoning model that publishes no reasoning_options gets low/medium/high, verified against the live endpoint — omp's bundled catalog would otherwise inherit a ladder Kenari rejects (minimal for mimo-v2-5 returns HTTP 400).
• If ~/.omp/agent/models.yml also declares a kenari provider, the two definitions collide — a config-sourced provider shadows an extension one. Keep the definition in exactly one place.

Scripts

Script	Purpose
bun run dev	Load this checkout into one throwaway CLI run (omp models ls -e) and list what it registers.
bun run typecheck	tsc --noEmit under strict.
bun run test	Alias of typecheck; there are no unit tests.
bun run smoke	Live end-to-end verification; needs KENARI_API_KEY and network.
bun run verify	typecheck then smoke — the full pre-publish gate.
bun run publish:dry	npm publish --dry-run: build the tarball, run the prepublish gate, send nothing.

prepublishOnly runs typecheck, so a broken tree cannot be published.

Releasing

bun run verify                 # 1. gate: typecheck + live smoke
bun run publish:dry            # 2. inspect tarball, confirm the gate fires
npm version patch              # 3. bump version + tag
npm publish --access public    # 4. upload

Verify

bun run verify                 # typecheck + live smoke
KENARI_API_KEY=... bun run smoke

# Or drive it straight from the CLI, loading just this checkout:
bun run dev

test/smoke.ts loads the plugin through omp's own extension loader and registry, then exercises the catalogue, a real chat completion, a real judgment, all four tools and the /kenari command. It needs the network and a credential, so it is deliberately not named *.test.ts and never runs under bun test.