nc://kilo · Nodecode

nodecode-kilo

A Nodecode cell for Kilo Gateway, a gateway that serves some six hundred models from many vendors behind one account. With it, /kilo login signs you in with a device code, /connect offers Kilo Gateway for a key, /models lists its models under their vendors' ids (anthropic/claude-opus-4.7, qwen/qwen3.7-max, ...), and a turn on kilo/<model> goes to https://api.kilo.ai/api/gateway over the OpenAI chat wire, with one quirk carried from omp: a Qwen model is told whether to think with enable_thinking, the Qwen dialect, instead of reasoning_effort.

Ported from oh-my-pi's kilo provider. See NOTICE.

Sign in

/kilo login

omp's Kilo login is a custom one: a device flow against Kilo's own device-auth endpoint. The command asks Kilo for a code and answers with the page Kilo names and the code. Open the page on any device, approve the code, and this machine notices within five seconds (it asks every five seconds until Kilo's code expires). A notice says when you are signed in or why the sign-in failed: denied, expired, timed out. Nothing listens on this machine.

The token goes into Nodecode's shared auth.json under oauth_tokens.kilo:

{"provider": "kilo", "access_token": "...", "refresh_token": "", "expires_at": 1791234567}

Kilo hands back a bearer with no refresh token, and omp keeps it for a year. So does this cell. After that, a turn is refused with "sign in again with /kilo login", and that notice stands until you do. Every other field of auth.json is kept, and the file stays mode 0600. /kilo status says whether you are signed in, and /kilo logout takes the entry out.

Key

A key works as well. Save it with /connect, or set KILO_API_KEY. The order: a key /connect saved (the core reads it before any cell is asked), then the sign-in, then KILO_API_KEY, as omp puts a stored credential before the environment. With none of them, nothing is sent as a key. The chat family's OPENAI_API_KEY never reaches Kilo.

Install

Copy this folder into ~/.nodecode/cells/nodecode-kilo/ and restart Nodecode, or run (restart-cells).

Configure

Optional. The section is on unless it says "enabled": false:

{
  "kilo": {
    // where Kilo Gateway is served; a relay's address goes here
    "base_url": "https://api.kilo.ai/api/gateway"
  }
}

Thinking

omp resolves each model's thinking dialect from its identity class. The Qwen class (every id naming qwen or qwq, and prism-ml/ternary-bonsai-2-27b) speaks the Qwen dialect: for a reasoning model, an effort sends enable_thinking: true and no reasoning_effort, and off sends enable_thinking: false. With no effort set, nothing is sent, so the provider's default applies, as the chat lane does for every provider. Every other model keeps the chat lane's reasoning_effort.

Not ported

Per-model replay rules omp keeps for a few models (thinking-requires-effort, reasoning content on every assistant turn for some DeepSeek ids). The chat lane already replays reasoning_content once a model streamed one, and treats every DeepSeek id as a thinking model.
Live model discovery. The core lists any credentialed catalog provider from <base>/models, which is what omp's Kilo discovery does.

Gaps

Per-model idle deadline. omp gives moonshotai/kimi-k2.6 a 300 s stream-idle timeout. A round's idle deadline is the frozen EFFECTIVE-PROVIDER-CONFIG-REQUEST-TIMEOUT (from providers.kilo.stream_idle_timeout_ms in the operator's config, for the whole provider), and NLE::WALK-PROVIDER-STREAM takes no keyword that would set it for one round. The seam would be a :request-timeout keyword on WALK-PROVIDER-STREAM, or a per-model rung in SNAPSHOT-EFFECTIVE-PROVIDER-CONFIG.

Test

From this repository's root:

tools/test-cell.sh nodecode-kilo

Kilo's device endpoints and every round are stubbed dex:post and dex:get calls, and the wait between two polls is stubbed out, so nothing reaches Kilo and nothing sleeps.

MIT licensed.