skip to content
$registry / workers

judge-typesafe

v0.2.5

TypeSafe evaluation (judge-typesafe::evaluate), model listing (judge-typesafe::models::list) and caller-scoped cancellation (judge-typesafe::cancel), with typed usage, provider errors, per-call retries and worker-owned credentials.

worker
iiiverified
1,879 installs473 in 7d83 today
install
$iii trigger compose::add worker=judge-typesafe
  • macOS: arm64 · x64
  • Linux: arm64 · armv7 · x64
  • Windows: arm64 · x64

exact versions are immutable; binary and bundle artifacts are digest-pinned.

configuration

iii-config.yaml
- api_key: null
  max_request_bytes: 8388608
  max_response_bytes: 8388608
  max_timeout_ms: 300000
  model: jev-latest
README.md

judge-typesafe

TypeSafe JEV provider for the judge hub. It turns JSON state and Noul, Choice or Score questions into typed, validated answers, sharing TypeSafe credentials and HTTP capacity across callers, with model listing and caller-owned cancellation. Callers go through judge::evaluate; the judge-typesafe::* functions below are the same contract addressed directly. They are registered as internal, so default discovery (engine::functions::list, directory::search_functions) shows only the hub.

Install

iii trigger compose::add worker=judge-typesafe

Compose starts the provider with its configuration dependency; installing judge pulls it in automatically. Use engine iii/v0.24.0-rc.2, the verified release with configuration::ensure used by this repository's CI. Wait for JEV's functions to register before making a call.

Quickstart

In the Console, open Settings → Workers → Judge TypeSafe and set API key (the form checks the saved key against judge-typesafe::models::list and offers the answered catalog in the Default model select), preferably as a secret://TYPESAFE_API_KEY reference to the secrets worker, or supply TYPESAFE_API_KEY in the JEV service's environment before starting it. See Configuration for precedence and reload behavior.

Ask whether a support ticket needs urgent attention:

iii trigger judge-typesafe::evaluate --timeout-ms 65000 --json '{
  "timeout_ms": 60000,
  "evaluations": [{
    "id": "ticket-42",
    "state": {"ticket": "Production checkout is down for all customers."},
    "questions": {
      "urgent": {
        "type": "noul",
        "instructions": "Does this describe an urgent production outage?",
        "criteria": {"true": "A critical production workflow is blocked."}
      }
    }
  }]
}'

A successful response looks like this; scores, usage and timing are illustrative:

{
  "status": "ok", "model": "jev-1.13.0",
  "results": {"ticket-42": {
    "answers": {"urgent": {"type": "noul", "noul": 0.98}},
    "usage": {"input_tokens": 120, "output_tokens": 8}
  }},
  "stats": {"attempts": 1, "requests": 1, "questions": 1,
    "input_tokens": 120, "output_tokens": 8, "elapsed_ms": 250,
    "usage_complete": true}
}

Check status before reading answers. A Noul value is between 0 and 1; choose your own decision threshold. Errors return status: "error" and a code with no partial results; a bus invocation failure is handled separately. The payload budget covers the whole call, including queueing and retries; the longer CLI timeout leaves room for transport overhead.

See the hub's mixed Noul/Choice/Score example, result and error handling, model listing and cancellation. Through the hub, request_id arrives here as :/, which is why this worker accepts ids of up to 512 bytes (judge_contract::MAX_PROVIDER_REQUEST_ID_BYTES).

Configuration

The Console entry defaults to judge-typesafe; set III_CONFIG_NAME=judge-typesafe-prod in the worker environment to use judge-typesafe-prod. The form masks the API key and exposes these defaults:

api_key: null                  # Fall back to the worker's TYPESAFE_API_KEY; or secret://NAME.
model: jev-latest               # Default unless a call supplies its own model.
max_request_bytes: 8388608      # Maximum JSON bytes per upstream evaluation.
max_response_bytes: 8388608     # Maximum bytes per upstream response.
max_timeout_ms: 300000          # Whole-call timeout ceiling in milliseconds.

A nonblank configured key overrides the worker's environment key. Clearing it restores the fallback. All configuration fields hot-reload for new calls; in-flight calls keep their snapshot, and rejected reloads keep the last valid value. Changing the process environment requires a restart.

The configuration service persists values at ./config/.yaml with its default filesystem adapter. Keep literal API keys out of committed project configuration: store the key in the secrets worker, list judge-typesafe in its consumers, and set api_key: secret://TYPESAFE_API_KEY (NAME matches ^[A-Za-z_][A-Za-z0-9_.-]{0,127}$). The reference is resolved through secrets::resolve at call time, never sent upstream as a key, and kept only in memory. It is re-read when secrets::changed names it (bound at boot; the engine holds the binding until a secrets worker registers the type), after 5 minutes, and 10 seconds after a failure. While the secrets worker is unreachable, a key already resolved keeps being used. A reference that does not resolve returns missing_key with provider_error.message naming the fix (secret NAME not found in the secrets worker, judge-typesafe is not allowed to read secret NAME; add judge-typesafe to the secret's consumers, secrets worker is not running) and never falls back to TYPESAFE_API_KEY. The reference itself is not secret; never paste a key after secret://. Any other scheme:// value, such as env://TYPESAFE_API_KEY, is not sent as a literal key either: it returns missing_key with api_key uses an unsupported env:// reference; use secret://NAME, which echoes only the scheme, and also never falls back. The form retains unknown values when editing other fields and shows errors returned by the configuration service; a ${TYPESAFE_API_KEY} value expands in the configuration service's environment instead of this process.

Consumer compatibility

Callers use the shared judge-contract request and response types. See shared-contract compatibility when updating existing consumers. JEV supplies evaluations; each caller owns its thresholds, ranking decisions and fallback policy.

For the full API, read the hub's reference.md. For source builds, UI work and local test commands, read CONTRIBUTING.md.