nc://kimi-code-delegate

Kimi Code Delegate: ask the locally signed-in Kimi Code CLI for read-only advice, as one tool the model calls. A Nodecode rewrite of s-a-s-k-i-a/hermes-kimi-code by s-a-s-k-i-a.

The agent doing the work stays the agent doing the work: this cell hands a second, different model one self-contained question and brings back what it says. Kimi is not a model route here — it never replaces this session's model, never takes a turn, and cannot be chosen as a fallback — and the call is read-only by construction:

• one fresh kimi acp process and one ACP session per call, then it is gone;
• the CLI's own kimi login is the credential, and the child is handed an environment narrowed to the names a CLI needs, so no other key of this process reaches it;
• a permission request is answered with the agent's own reject_once option, and a filesystem or terminal request with a JSON-RPC error;
• an empty answer is a failure, not a turn: the cell says the provider status is unknown rather than guessing quota.

Get it

Tell the agent "add the kimi-code-delegate cell", or run nodecode add nc://kimi-code-delegate.

Set it up

Install the Kimi Code CLI and sign in once, in a terminal:

kimi login

The cell finds kimi on PATH, or the CLI's own ~/.kimi-code/bin/kimi. Nothing else is needed; there is no key of its own. The section is optional and lives in ~/.nodecode/config.jsonc (/settings, or the file; config.example.jsonc has the shape):

"kimi-code-delegate": {
  "command": "kimi",                 // a name on PATH, or a path
  "working_directory": "",           // where a call runs; empty = Nodecode's own
  "model": "k3",                     // or "k3-256k"
  "thinking": "auto",                // or an option id the CLI offers
  "timeout_seconds": 300,
  "startup_timeout_seconds": 10
}

Check the wiring before spending a real call: /kimi-code doctor runs the handshake alone and says who answered, the version, the models the CLI offers and its current effort.

Use it

The model calls kimi_code_delegate with a prompt — work it cannot see any of this conversation, so the prompt must stand alone — and optionally a model (k3, k3-256k) and a thinking effort. The answer is JSON: the agent's text, the model and effort it ran on, and delegate_only.

Ask for it in words and the agent decides; ask explicitly when you want the second opinion:

Ask Kimi to review this migration plan for anything that would break on rollback.

It is most useful where a second model is worth a call: a design about to be committed to, a review of a diff or a plan, a second reading of a rule. It cannot edit files, run commands or answer with tool calls, so a task that needs any of that is the worst fit.

/kimi-code with anything but doctor prints its usage.