nodecode-jev

Jev: fast advisory judgments -- triage a request, pick a next step, rate relevance -- answered as probabilities. A Nodecode rewrite of ourines/hermes-jev by ourines.

Ask one bounded question and get an answer you can act on without a second model call:

(jev:triage "The export button writes an empty CSV on the staging build; no logs yet.")
;; => (:CATEGORY (:TYPE :CHOICE :CHOICE "bug" :PROBABILITIES (("bug" . 0.6) ...) :CONFIDENCE 0.55)
;;     :MISSING-CONTEXT (:TYPE :BOOL :BOOL 0.2)
;;     :REQUIRES-APPROVAL (:TYPE :BOOL :BOOL 0.8)),
;;    (:ENABLED NIL :THRESHOLD NIL :REVIEW-QUESTIONS () ...),
;;    "typesafe/jev-latest"

Jev answers typed questions -- a choice among labelled options, a yes/no, a rubric score -- about one state, and nothing else: no conversation, no tool call, no action. The model decides; Jev only says how likely. Its answer is explicitly advisory: it grants no approval and executes nothing.

Verb	What it answers
`(jev:triage state)`	what kind of work this is (bug, feature, research, docs, operations, unknown), whether context is missing, whether it needs authorization
`(jev:next-step state)`	continue, retry, change_approach, ask_user, escalate or stop -- and whether the attempts are stuck
`(jev:relevance state)`	how directly an item helps a goal, 0 to 2, and whether the goal or the item is missing
`(jev:ask state :preset :relevance)`	the same three presets, by name
`(jev:ask state :questions ...)`	your own questions, in one call
`(jev:presets)`	the three rubrics, offline
`(jev:status)`	backend, model, where the key comes from; no network

Every judgment also answers a review value: with a threshold set in the section, the answers whose probability sits near the middle, whose confidence is low, that come back unknown, ask_user or escalate, or that say approval is likely, are named as worth a human look. It is a heuristic, not a calibration, and it never replaces a safety check.

Get it

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

Set it up

Pick one backend, as the original insists -- there is no fallback from one to the other:

typesafe (default model jev-latest). Make a key at https://console.typesafe.ai/, then set TYPESAFE_API_KEY, or save it in ~/.nodecode/auth.json under api_keys.typesafe the way /connect keeps a key. The nodecode-typesafe cell reads the same entry, so one key serves both.
cloudflare (default model typesafe/jev, the same model through AI Gateway). Set CLOUDFLARE_JEV_API_TOKEN -- a Workers AI token -- or save api_keys.cloudflare-jev, and put the 32-hex account id in the section. An authenticated gateway with Unified Billing credits is what makes the route run; requests ask the gateway to keep no log, use no cache and allow one attempt.

{
  "jev": {
    "backend": "typesafe",          // or "cloudflare"
    "account_id": "",              // cloudflare: 32 hex digits
    "gateway_id": "",              // cloudflare, optional
    "model": "",                   // empty: jev-latest / typesafe/jev
    "review_threshold": ""         // empty: heuristic off; else 0.5 < n < 1
  }
}

(jev:status) and /jev say what is configured without asking anyone. A request that fails says why in one line -- no key, an HTTP status, or Cloudflare's numeric provider code -- and never echoes the key.

Use it

In a turn, put the observations in state -- the goal, the item, the last attempts, what you know -- and read the answer with getf:

(multiple-value-bind (answers review backend)
    (jev:next-step '(:goal "ship the migration guide"
                     :attempts ("draft written" "linter keeps failing on headings") 
                     :permissions "edit docs, no deploy"))
  (getf (getf answers :next-step) :choice)   ; => "retry"
  (getf review :review-questions))           ; => () with no threshold set

A question of your own, with three primitives in one paid call:

(jev:ask "The patch renames the public function parse-config."
         :questions '(:risky (:type "bool" :instructions "Does this change risk breaking callers?"
                              :criteria (:true "an API changes" :false "internal only"))
                      :kind  (:type "choice" :instructions "What kind of change is this?"
                              :criteria (:fix nil :feature "adds" :refactor "restructures"))
                      :size  (:type "score" :instructions "How large is the change?"
                              :criteria ("trivial" "small" "medium" "large"))))

A choice needs two or more options, each a rubric or nil; a bool's criteria are optional; a score needs two or more levels, lowest first. Give exactly one of :preset and :questions. An answer that is missing, of the wrong type, an option that was never offered or a level outside the rubric is a refusal naming the question -- never a guess.

(help :jev) carries the same manual while the cell runs.

Test

From this repository's root:

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

Not ported

The Hermes plugin surfaces: the interactive hermes jev setup wizard, its status/test/guide/ evaluate --file subcommands and its JSON request files (a call through eval is the request), and the two bundled skills (the manual the model reads replaces them). No retries and no cross-backend fallback, as the original. See NOTICE.

MIT licensed.