pi-jev-typesafe

TypeSafe Jev inside Pi. Jev is TypeSafe's System One judgment model: send it some state and typed questions, get calibrated probabilities back in about a second, for a fraction of a cent. This package gives Pi one thing done carefully:

• jev_ask tool — batched Choice / Score / Noul evaluation with question linting that catches the failure modes the TypeSafe docs warn about before your tokens are spent, model discovery (/jev models), per-request model pinning, and probability-bar rendering in the terminal.
• /jev command — enable · disable · status · test · models.
• Budget rails — per-session attempt cap, optional daily caps (requests / input tokens / USD), persisted across restarts, checked before a request is submitted.

Independent project. Not affiliated with TypeSafe AI or the Pi authors.

Install

pi install npm:pi-jev-typesafe
# or from git:
pi install git:github.com/nardinmarcus/pi-jev-typesafe@v0.1.0
# or from a local checkout:
pi install /path/to/pi-jev-typesafe

Requires Pi 0.85+ and Node.js 22+. Zero runtime dependencies — pi bundles everything the extension imports (typebox, pi-ai, pi-tui).

Setup

1. Get a key at console.typesafe.ai.
2. export TYPESAFE_API_KEY=... in your environment (~/.zshrc works; the extension also reads it as a fallback if the env var is absent from the pi process).
3. Inside Pi, run /jev enable and confirm the data notice — or set PI_TYPESAFE_JEV_ENABLED=1 for headless runs.

/jev status shows the key source, opt-in state, session attempt count, today's persisted usage, estimated spend, and active caps.

The three question types

Type	Asks	Criteria	Returns
choice	Which of these options fits?	options map { label: description | null } — include a no-match option	chosen option, probability per option, confidence
score	Where on this ordered rubric?	array of ≥ 2 concrete level descriptions	position (may be fractional), probabilities, confidence
noul	Is this true?	optional { true, false } descriptions	P(yes)

Questions run in parallel and cannot see each other. Put each item in a named state field (reports.r1) and ask one narrow judgment per question per dimension.

Question linting

Before submitting, jev_ask checks your questions and appends non-blocking warnings to the result:

• a choice with no no-match option (other, unclear, 其他…) — the model cannot pick an option you omitted;
• score levels that are bare degree words (low / medium / high / 中等 / 一般) instead of concrete situations;
• instructions so short they cannot carry the full judgment.

Question text is the whole program: Jev answers exactly what is asked, and ambiguity shows up as a middling probability rather than an error.

Budget and safety

• Per session: 20 attempts by default (PI_TYPESAFE_JEV_MAX_REQUESTS), reset on session start/reload.
• Per day (persisted in ~/.pi/agent/pi-jev-typesafe/usage.json, rolls over at local midnight):
  • PI_TYPESAFE_JEV_MAX_REQUESTS_PER_DAY
  • PI_TYPESAFE_JEV_MAX_INPUT_TOKENS_PER_DAY
  • PI_TYPESAFE_JEV_MAX_USD_PER_DAY (input tokens only — output is free)
• 20s timeout, one retry on transient faults (429 / 5xx / network), response shape validated before you see it.
• Only the submitted state and questions go to api.typesafe.ai. Nothing else is collected. Error messages never contain upstream response bodies, headers, keys, or your submitted content. The API key is never logged.

Model selection

model defaults to jev-latest. /jev models lists what your account can use.

• jev-latest — most recent stable release (currently jev-1.13.0)
• jev-preview — may move ahead of stable
• pinned ids (jev-1.13.0) — use when confidence thresholds were tuned against a specific version

Library use

Internal helpers are exported for reuse and testing:

import { validateRequest, lintRequest, prepareArguments, formatResult } from "pi-jev-typesafe";

Development

npm install
npm run check        # typecheck + tests
npm run build        # emit extensions/index.js (committed: git installs do not build)

License

MIT