jevvy

jevvy is a Pi extension that uses Jev to review comments, functions and tests. Each pack asks fixed questions and returns the answers with their source evidence. Pi can inspect that evidence before recommending changes. Jevvy scans leave source files unchanged.

Source parsing supports JavaScript/JSX, TypeScript/TSX, Rust, Python and Solidity. The Functions and Tests packs adapt Kiln Code Savers rubrics for unnecessary complexity and tests of application-owned behavior. They also include narrow checks developed through live Jev experiments.

Install

Requires Pi and Node 26 on macOS or glibc Linux, ARM64 or x64. See tested environments for platform coverage.

pi install npm:jevvy
cd /path/to/your/project
pi

Restart an existing Pi session after installation. Add -l to the install command to install only for the current project.

To install from Git instead, use pi install git:github.com/timbrinded/jevvy.

In Pi, preview the comments affected by your working changes:

/jevvy comments --working --dry-run

A dry-run shows the selected comments and planned requests without API calls or an API key. For live analysis, set TYPESAFE_API_KEY in your shell before starting Pi, then omit --dry-run. Requests use your TypeSafe account.

Scan comments

File paths are relative to Pi's working directory. Working scans compare against HEAD and include non-ignored untracked files. Branch scans compare the base/head merge-base with head; replace origin/master with your base branch.

Diff scans include unchanged comments attached to changed code. A clean working tree has no changes to review; use --files to scan existing files. Add --dry-run to any scan to preview it, or use /jevvy cancel to stop a running scan.

Scan functions and tests

/jevvy functions --files src/orders.ts
/jevvy tests --files test/orders.test.ts --context-files src/orders.ts package.json

Both packs also support --working, --base <ref> [--head <ref>], and --dry-run. Supporting files are explicit: include the relevant implementation, manifest and contract so a test can be assessed against the behavior it protects. They are captured from the selected snapshot and subject to the context budget. Missing evidence remains visible; Jevvy does not infer a complete call graph or test suite. See pack rubrics and limits.

Read results

Each scan saves a bundle containing the assessments and the source used to make them. The result card shows coverage, errors and the bundle ID. Completed answers remain available after cancellation or failures in other requests.

Open the inspector to browse selected units, saved source and question definitions:

/jevvy inspect <bundle-id>

Left and right change pages, up and down scroll, and Tab switches views. Escape or your configured Pi cancel key closes the inspector.

Retrieve an overview or a page of comments:

/jevvy results <bundle-id>
/jevvy results <bundle-id> --view units --limit 5

Pi can also use jevvy_comments, jevvy_functions, jevvy_tests, and jevvy_results. Reading saved results makes no API calls. See the sample report for the output format; its values are synthetic.

For a code-pack result, select a question and rank its issue outcome:

/jevvy results <bundle-id> --view units --labels owned_behavior --sort owned_behavior --outcome issue --min-probability 0.8 --min-confidence 0.7 --include-context --include-definitions

Those cutoffs are example review settings, not calibrated accuracy guarantees. Filters operate on saved answers. The full distributions, unknown outcomes and unfiltered results remain available. A scan that raises no candidate does not establish that the code or tests are correct.

Data and settings

Live scans send selected units and their bound context to TypeSafe. Context can include an entire small file. Reports and retrieved code enter Pi's model conversation. Bundles store full captured files locally so results retain the source as it was when scanned.

Bundles and cached answers are stored in ~/.cache/jevvy. Set JEVVY_STORAGE_DIR to change this location; keep it outside tracked project files. Saving a new bundle removes results and cached answers older than 30 days by default.

For other settings, set JEVVY_CONFIG to a JSON object that overrides the configuration defaults.

Development

Use the Node version in .node-version and the pnpm version in package.json. TypeScript 7 checks types and builds the library; Node runs the tests directly.

git clone https://github.com/timbrinded/jevvy.git
cd jevvy
pnpm install --frozen-lockfile
pnpm run check
pnpm run smoke
pnpm exec pi install .

check runs formatting, lint, typecheck, tests and build. smoke checks Pi integration. Neither makes paid API calls. Use pnpm run smoke:live with TYPESAFE_API_KEY for a live check.

The last command registers the checkout as a local Pi package. Restart Pi after source edits.

• Release guide: Pi discovery, npm packaging and publication.
• Functions and Tests: rubrics, evidence requirements and Kiln attribution.
• Verification: test coverage and detailed checks.
• Bundle contract: schemas, validation and retrieval rules.

License

Apache-2.0. Separate third-party notices apply to the Solidity grammar and adapted Codesavers material.