nc://decisions

Decisions: ask yes/no questions about a text and get, for each, the probability that it is true, from an OpenAI-shaped Decisions endpoint. A Nodecode rewrite of eve-ai-dev/hermes-decisions-plugin by eve-ai-dev.

A model reads sentences that are not quite empty — a ticket that may be urgent, a paragraph that may be a complaint, a plan that may be out of scope — and asks a decision endpoint for a number instead of an opinion. Two tools do it, and the answer is the endpoint's own JSON, extra members included.

Get it

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

Set it up

A key is required: it goes in the decisions section of ~/.nodecode/config.jsonc, either written as api_key (kept out of the file, in ~/.nodecode/secrets/) or named by api_key_env, the variable that holds it — OPENAI_API_KEY by default.

{
  "decisions": {
    "api_key_env": "OPENAI_API_KEY",
    "endpoint": "https://api.openai.com/v1/decisions",  // any absolute http(s) URL, no query
    "model": "gpt-6-luna",                              // whatever the endpoint accepts
    "timeout_seconds": 120                              // 1..600; a call past it is not retried
  }
}

An endpoint of your own needs endpoint and model changed together; the key goes to that endpoint and nowhere else. Without a key the cell still starts, and each call says which setting fills it. /settings and the Settings tab draw the same five fields, and (restart-cells "nodecode-decisions") picks up an edit without a relaunch.

Use it

While the cell runs, (help :decisions) teaches both tools, and every request carries one line naming them.

• decisions_submit takes input (the text, at most 256 KiB with its questions) and questions, one to 64 of them, each {"type": "predicate", "instructions": "is it urgent?"}. Ask a question in the affirmative: the number is the probability that it is true.
• decisions_status takes job_id and answers that submission of this session: running, or completed with the reply object, or failed with the reason.

background defaults to true, as in the original: the call answers a job id at once and leaves the request running on a thread of its own, so a slow endpoint never holds a turn. background: false waits for the probabilities and answers them, and the job is still there afterwards to read back. /decisions lists this session's submissions and what became of them, and /decisions <job_id> answers one job's whole record.

An answer looks like {"type": "predicate", "name": null, "probability": 0.87}. Four submissions run at once and 32 finished ones are kept, all in memory: a restart keeps none, and a job belongs to the session that started it. Nothing is retried — a failed call says whether the outcome is uncertain, and asking again is a second call.

The cell refuses what the endpoint cannot mean: a reply with a different number of answers than questions asked, a probability outside 0..1, a non-predicate or named answer, a body over 1 MiB, and a URL with a scheme, query or fragment it cannot POST to.

Tests

.github/test-cell.sh tools/nodecode-decisions

Offline: the one network call is bound to a scripted reply and a background job runs on the test's own thread, so the request body, the job record, the session's ownership of a job and each refusal are read as facts. Eleven tests, including the cell's start and stop, both tools, the slash command, and a section with no key.