nodecode-jev-routing
Jev Routing: per-prompt model routing from TypeSafe Jev. A Nodecode rewrite of apoapostolov/hermes-agent-awesome-plugins by apoapostolov.
Type normally. When a prompt is worth thinking about, this cell asks TypeSafe Jev four typed questions about it -- what kind of work it is, how hard it is, how much model capability it deserves, and whether it needs deep reasoning -- and composes the answers into a capability tier. Then it names the model your own routing table holds for that tier: the fast free model for a one-line fix, the frontier model for an architectural change.
The original does this inside every Hermes turn and rewrites the request.
Nodecode hands a cell no per-turn model hook, so this one is the decision
engine itself: the model calls it through eval and acts on the answer --
run the subtask on that model, switch with /model, or just know what the
turn is worth.
Get it
Tell the agent "add the jev-routing cell", or run nodecode add nc://nodecode-jev-routing.
Set it up
Two things: a TypeSafe key, and a routing table.
Make a key at https://console.typesafe.ai/. Set TYPESAFE_API_KEY, or keep
it in ~/.nodecode/auth.json the way /connect keeps a provider key -- the
same api_keys.typesafe entry the typesafe cell reads, so one key serves
both:
{"api_keys": {"typesafe": {"provider": "typesafe", "key": "..."}}}
The routing table is a JSON file you write, pointed at by config_path:
{
"jev-routing": {
"config_path": "~/routing.json", // empty leaves the cell inert
"mode": "shadow", // shadow, notify, or off
"base_url": "", // empty: TYPESAFE_BASE_URL, else https://api.typesafe.ai
"model": "", // empty: TYPESAFE_DEFAULT_MODEL, else jev-latest
"timeout_ms": 3500
}
}
A base_url of your own speaks the same POST /v1/systemone (a local DJev
server, say); with no key saved or set, the request goes without
Authorization. /jev-routing says where the table is, which chains it holds and
what the ledger says, asking no one.
The routing table
{
"routes": {
"quick": [{"provider": "opencode-go", "model": "beam-free"}],
"standard": [{"provider": "opencode-go", "model": "glm-5.3"}],
"high": [{"provider": "openai-codex", "model": "gpt-6-sol", "thinkingLevel": "medium"}],
"premium": [{"provider": "openai-codex", "model": "gpt-6-sol"}],
"xpremium": []
},
"kindMinimumTier": {"plan": "high", "review": "high", "implement": "standard"},
"kindModels": {"implement": [{"provider": "openai-codex", "model": "gpt-6-sol", "minTier": "standard", "priority": 20}]},
"confidenceThreshold": 0.34,
"minPromptChars": 12,
"budget": {"dailyUsd": 10, "monthlyUsd": 150, "softRatio": 0.7, "hardRatio": 0.9},
"free": {"enabled": true, "policy": "fallback-only", "pool": [{"provider": "opencode-go", "model": "beam-free"}]},
"pricing": {"opencode-go/glm-5.3": {"input": 1.4, "output": 4.4}}
}
A tier the file leaves empty stays empty; the loader never invents a model
list. kindModels names a model per kind of work, from a tier up;
kindMinimumTier is a floor a protected kind never routes below. Over the
soft budget ratio the pick drops one tier, over the hard ratio it drops to
quick unless the demand is already premium, and an empty free pool changes
nothing. A pricing row is per million tokens, in dollars.
Use it
Through eval, while the cell runs:
(jev-routing:route "Refactor the parser into two modules.") ;; => (:TIER "premium" :DESIRED-TIER "premium" :PROVIDER "openai-codex" :MODEL "gpt-6-sol" ;; :THINKING NIL :KIND "implement" :DEMAND 3.0 :CONFIDENCE 0.81 ;; :BUDGET-PRESSURE 0.0 :LOW-CONFIDENCE NIL :DOWNGRADED NIL :NOTES NIL ;; :GLYPH "." :LINE "tier premium · openai-codex/gpt-6-sol · kind implement · ..." ;; :LATENCY-MS 96) (jev-routing:route "ok") ; => (:SKIPPED "acknowledgement") -- nothing sent (jev-routing:judge "Review this diff.") ; the four answers alone, for the why (jev-routing:spend) ; today, this month, the caps, the pressure (jev-routing:record "gpt-6-sol" :provider "openai-codex" :input 12000 :output 800)
HISTORY (the recent conversation, at most 4000 characters go),
ACTIVE-MODEL and CONTEXT-TOKENS are optional, and they make a short
follow-up route on what it means rather than on its eleven characters.
/jev-routing says where the table is, its chains and the ledger. /jev-routing route TEXT
names the tier a text deserves; /jev-routing why re-asks Jev about the last prompt
and shows the full judgment, probabilities and all; /jev-routing spend is the
ledger. (help :jev-routing) answers this manual.
What leaves the machine. The prompt, up to 4000 characters of the
conversation you pass as history, the working directory, the active model,
an estimate of the context in use, and the spend snapshot, to TypeSafe alone
(https://api.typesafe.ai unless you set base_url). Nothing else, and
nothing at all for a prompt that does not re-route. A missing key, a timeout
or a bad answer signals JEV-ERROR and changes nothing.
The ledger is <table-stem>-state.json beside the routing table: days
and months, split by model, with the original's shape. A reported USD total
wins; otherwise your pricing rows price the tokens; an unpriced model
records tokens and no dollars rather than inventing spend. Only what you
record is in it -- no turn is recorded for you.
Test
From this repository's root:
.github/test-cell.sh tools/nodecode-jev-routing
Not ported
- The turn. The original rewrites the provider request per turn, offers an approve row for guarded tiers, and moves the session through the app's model switch. A cell has no per-turn model hook, so this one answers and the caller acts.
- The catalogue, quota and cache hold.
decide.pyfilters a live model list, reads Codex quota windows and holds a warm cache. A cell reads no catalogue, so every configured row is eligible -- the catalog edition's own behaviour -- and no pick is held for the cache. /jev-routing suggest, the ranking and the generated layer, and the desktop half in JavaScript.- Confirm clicks, pinning and stickiness (
/jev-routing revert, thejev-confirm:marker,stickiness), which belong to a turn stream. TYPESAFE_API_URL: a base is set with the section'sbase_urlorTYPESAFE_BASE_URL.
