@jev-harness/cli

Two binaries for TypeSafe Jev — the System One decision model:

• jev — the unified command line: ask, models, eval. Debug question design without building an adapter, and run evaluations without remembering package names.
• jev-gate — a small CI/subagent gate around one Jev judgment over a git diff, file, or stdin.

Install

npm install -g @jev-harness/cli   # or: npx @jev-harness/cli <command>

From the monorepo:

npm install
npm run build --workspace=@jev-harness/cli

Requires TYPESAFE_API_KEY in the environment (TYPESAFE_BASE_URL, TYPESAFE_DEFAULT_MODEL, JEV_TIMEOUT_MS are honored too).

jev — commands

jev ask — one batched ask

# inline questions, default {} state
jev ask --questions '{"gate":{"type":"noul","instructions":"destructive?"}}'

# state and questions from files
jev ask --state state.json --questions questions.json

# envelope on stdin: {"state": ..., "questions": ...}
cat request.json | jev ask

# '-' reads that input from stdin
jev ask --state - --questions questions.json < state.json

Prints the full JSON response (answers with calibrated probabilities per question, plus usage) to stdout — pipe it into jq or a file. --model, --base-url, and --timeout-ms override the environment for one call.

jev models

jev models   # → ["jev-latest", ...]

jev eval

Delegates to jev-eval — same flags:

jev eval --dataset cases.jsonl --out report.json

jev-gate — usage

Judge a diff:

jev-gate --diff -c "all exported functions are documented"

Judge a file or piped output:

jev-gate --file ./report.txt --criteria "the report contains no failures"
npm test 2>&1 | jev-gate "zero test failures" --json

Exit codes for jev-gate:

• 0: Jev probability meets the threshold.
• 1: Jev judged the criteria below the threshold.
• 2: the gate could not evaluate (missing key, invalid arguments, unreadable file, or API error).

Use --fail-open only when the surrounding pipeline explicitly prefers availability over enforcement. It reports the error but exits 0.

Exit codes (jev)

| Code | Meaning | | ---- | ------------------------------------------------------ | | 0 | Success | | 1 | Runtime error (missing key, API failure) | | 2 | Usage error (bad flags, invalid JSON, unknown command) |

Programmatic use

import { runCli } from "@jev-harness/cli";

const code = await runCli(["ask", "--questions", "..."], {
  fetchImpl: mockFetch,   // injectable for tests
  stdin: "",              // pre-read stdin
  out: (s) => ...,        // output sinks
  err: (s) => ...,
});

Scope

This is a semantic acceptance check, not a replacement for tests, linters, typecheckers, or security scanners. The state is sent to TypeSafe's hosted Jev API; do not pipe secrets or credentials into it.

License

Apache-2.0.