nc://orquehelx

OrqueHelx: sends each subagent to the subscription you pick, one route per provider and model, and says plainly when one runs out of quota. A Nodecode rewrite of TicoraX/OrqueHelx by TicoraX.

Get it

Tell the agent "add the orquehelx cell", or run nodecode add nc://orquehelx.

Set it up

No key of its own: a route names a provider you already have configured, so the subagent runs on credentials that are already there. Write the routes in the orquehelx section of ~/.nodecode/config.jsonc (/settings, or the file; config.example.jsonc has the shape):

"orquehelx": {
  "policy": "ask",
  "seconds": 900,
  "routes": [
    { "name": "claude", "provider": "anthropic", "model": "claude-haiku-4-5" },
    { "name": "codex",  "provider": "openai-codex", "model": "gpt-5.5" }
  ]
}

A malformed route — no name, a name that is not lowercase letters, digits or -, no provider, no model, an unknown field — is skipped rather than guessed at, and a standing notice says which and why; a repeated name keeps the first. The section is read when the cell starts, so (restart-cells "nodecode-orquehelx") is what a rewritten route list needs.

Use it

Ask for a task to go to a route. The agent calls delegate_to with the route's name, the goal, and any context the subagent cannot see:

delegate_to(route: "codex",
            goal: "Find every caller of parse_header and say which ones pass a null buffer.",
            context: "src/parse.lisp, the header parser")

• The goal runs in a fresh subagent session pinned to the route's provider and model: its own transcript, the same folder, listed on /sessions under the session that asked. It never sees the parent's conversation.
• The answer is JSON: status, the route it ran on, the child's summary, and its session id.
• Only route names and the model behind each are in the tool's schema — the provider and the endpoint are never tool parameters, so the agent cannot invent a route, and a name outside the list is refused before anything runs.

/ohx routes lists every route with its provider and model. /ohx quota is one line per route: a provider the cell has seen run out of quota, with the hour it was seen and the reset it named, or no data — a figure is never estimated. What happens after an exhaustion is the policy:

• ask: the answer says status: out_of_quota, names the reset it knows (or that it knows none) and the other routes, and tells the parent to ask you and wait. It cannot delegate again in that turn until your answer opens a new one; subagents already running still finish.
• parent_decides: the parent may resend the task on another route whose model suits it, and says which one it chose and why.
• wait: the parent tells you and does nothing else with the task.

A route counts as out of quota only when the provider's own words say so (a 429 that speaks of a usage limit or a quota, or a RESOURCE_EXHAUSTED), never on a plain failure.

What is not here

The measured usage windows Hermes has (fetch_account_usage, a percentage and a reset time per provider): Nodecode has no such meter, so an exhaustion is recorded from a delegated subagent's failure and /ohx quota prints no percentage. Also not here: the pre_tool_call hook that blocks Hermes's own delegate_task (a Nodecode subagent is launched by the cell itself), and the v0.1 Spanish config names.

License

MIT. The original plugin is Copyright (c) 2026 TicoraX: see LICENSE and NOTICE.