nc://ordaprompt

OrdaPrompt: routes an ask to a kept topic or a live session, and abstains when it cannot tell. A Nodecode rewrite of aska-digital/protean-ordaprompt by aska-digital.

The original is a fail-closed comparative classifier for request routing. Given a request and a candidate set, it scores every candidate in one batch -- the kept topics, a novel sentinel, an ambiguous sentinel, the live sessions and a new_session sentinel -- and answers one of three bands:

Abstention is a first-class answer, not a failure. Session reuse is stricter than topic assignment: a session that wins on similarity but carries a contamination flag, or is near-tied with starting fresh, never routes automatically. Every decision is one call, and a hashes-only receipt can be kept of it.

In Hermes the whole surface was a command line, and its register() declared no tools, hooks or settings. Here it is a tool the model calls and a command the operator types; the band table above is the original's, line for line.

Get it

Tell the agent "add the ordaprompt cell", or run:

nodecode add nc://ordaprompt

Set it up

Nothing is required. The section's defaults are the original's thresholds, and the scoring backend is offline.

{
  "ordaprompt": {
    "calibration_active": true,
    "receipts_file": "~/.nodecode/ordaprompt/receipts.jsonl"
  }
}

calibration_active is the one switch worth knowing. The original shipped the automatic band locked off (calibration_model_id: "none") because its synthetic calibration was above the error gate it set for itself, so every answer was confirm or abstain. A locked band here is the same: with calibration_active false nothing routes on its own, every answer escalates or abstains, and the code says calibration_missing. The default is true, because this cell's backend has no calibration file to validate against; turn it off to get the release's posture back.

Thresholds are integers in hundredths -- the section has no float type -- so 85 is the original's tau_topic = 0.85. A member is set in /settings or in ~/.nodecode/config.jsonc, then (restart-cells "nodecode-ordaprompt").

Use it

The tool, from the model's side:

{
  "prompt": "the billing export writes an empty file on staging",
  "topics": ["billing", "deploy"],
  "sessions": [
    {"session_id": "s-3", "summary": "billing export rewrite", "tokens": 12000,
     "flags": []}
  ]
}

and the answer it reads back:

decision topic_assign, band automatic
topics: billing 1.00, ambiguous 0.00, deploy 0.00, novel 0.00
sessions: s-3 0.37, new_session 0.35
receipt: ~/.nodecode/ordaprompt/receipts.jsonl
bands: topic automatic; session abstain_or_new_session (below_margin)

The bands line is the decision's whole reason: the effective band is the more conservative of the two surfaces, and the code names which rule fired (calibration_missing, ambiguous_won, below_margin, contaminated_high, new_session_argmax, below_min_utility, below_min_score).

From eval, or from a layer:

(ordaprompt:route "the billing export is empty" '("billing" "deploy") nil)
;; => (:BAND "automatic" :DECISION "topic_assign" :CODE NIL ...)

The command:

/ordaprompt                                how it is configured, and how to call it
/ordaprompt billing,deploy :: the ask      one decision over two topic labels

(help :ordaprompt) carries the same in one paragraph.

The backend, and what changed hands

The original scored a prompt by hashing it into a 16-dimensional vector and comparing that with a hash of each candidate id: deterministic, offline, and with no relation to what the prompt actually said -- it was the fixture backend, waiting for a real NLI model behind the same interface. This cell scores by words instead: a candidate's score is the fraction of its own words the prompt carries (deploy meets deployment, but bill does not meet billing), novel is the residual, and ambiguous is how tied the top two are once something scores. Everything above the score -- both bands, the margins, the session utility, the sentinels, the effective band, the codes -- is the original's.

The receipt

With receipts_file set, every decision appends one JSON line: the prompt's sha256:, the candidate labels, the scores, the utilities, the band, the code and the router version. Never the prompt, never a file, never a path -- the original's receipt writer refused free text, and this one has nowhere to put any. A receipt that cannot be written says so in the answer rather than being lost.

Tests

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

13 tests, offline: the cell's start and stop, a named topic routing, an unmatched ask proposing a new topic, a locked band escalating, a call with no candidate refusing, a clean session reused, a contaminated one not, a receipt holding hashes and not words, and both command verbs.

Gaps

• The original's OpenRouter adapter (provisional taxonomy proposals, hashes only, disabled by default) is not carried: there is no key to set and no proposal to make, so a novel ask answers novel_propose and leaves the naming to the caller.
• confirm is not a decision of its own in the original either -- it is the escalate band -- and here the ranked list is the confirmation.
• The calibration document (--calibration, Platt scaling per surface) and the eval harness (eval/) are not carried: both existed to prove the automatic band's error gate on synthetic fixtures. calibration_active is the whole of what the gate decided.
• classify and route-session are one tool here: sessions are an argument, and leaving them out is the topic-only call.

Credit

The router, its bands, its thresholds, its sentinels, the contamination weights and the receipt shape are aska-digital's, from aska-digital/protean-ordaprompt at commit 60e89af, MIT. See NOTICE.