Jevort

A Pi extension that uses TypeSafe AI's Jev (jev-latest) to recommend a reasoning-effort tier for each user prompt. By default it only shows a notification; it never adds slash commands or changes the thinking level unless automatic application is explicitly enabled.

Requirements

• Node.js 22.19 or newer (required by Pi)
• Pi with extension support
• A TypeSafe API key (TYPESAFE_API_KEY)

Install

Install the published npm package for your user:

pi install npm:jevort

Or install the private/public GitHub repository directly:

pi install git:github.com/aklsh/jevort

Then set TYPESAFE_API_KEY in the environment that launches Pi. To try the extension from a local checkout:

npm install
export TYPESAFE_API_KEY="your-key"
pi --extension ./jevort.ts

Behavior

On Pi's before_agent_start event, the extension sends each non-empty prompt to TypeSafe and asks Jev to choose among:

• low — short, scoped task or tiny reversible edit
• medium — routine work with little architectural uncertainty
• high — complex debugging/design or meaningful integration risk
• xhigh — broad architecture, subtle correctness, or high stakes

A recommendation is accepted only when its selected-choice probability is at least 0.5; otherwise the classifier uses medium. The suggestion is adapted to Pi's common thinking levels using the active model's reasoning capability and thinkingLevelMap: unsupported levels are mapped to the nearest available tier. A model without reasoning support is left unchanged. If the TypeSafe request fails, the extension leaves the current level unchanged and reports a warning.

By default Jevort only notifies you of its suggestion; it does not change the active thinking level. Use Pi's built-in /thinking control to apply a suggestion manually. To control automatic application during a Pi session, use /jevort (equivalent to /jevort toggle), /jevort on, /jevort off, or /jevort status. This runtime setting resets when Pi restarts. To enable auto-application at startup, set JEVORT_AUTO_APPLY=1 (accepted values are 1, true, yes, and on) in Pi's environment. The extension checks the active model's reasoning support and thinkingLevelMap, maps unavailable recommendations to the nearest supported tier, and leaves the current setting unchanged when reasoning is unavailable.

Privacy and network behavior

The prompt text is sent to TypeSafe for classification. The API key is read from TYPESAFE_API_KEY. Requests use jev-latest, a 2.5-second timeout, and no retries. Do not use the extension for prompts you are not permitted to send to TypeSafe.

Development

npm run typecheck

The typecheck requires the dependencies from package.json to be installed. The extension relies on Pi's before_agent_start event and the model metadata exposed on its extension context. The package manifest declares jevort.ts as its Pi extension entrypoint and includes the TypeSafe SDK as a runtime dependency.