Jev Decision Gate by BPJ

What Jev Decision Gate by BPJ does

Jev Decision Gate by BPJ runs an issue-triage workflow against JSON data stored on the local machine. Each issue can be assessed for kind, affected module, and information sufficiency. The bundled taxonomy supports bug, feature, question, and unknown kinds; auth, api, ui, docs, and unknown modules; and either sufficient or missing information.

The workflow can use a local rules baseline or optionally send selected issue text to TypeSafe's Jev API. Results can be calibrated against human-reviewed examples and tested on a separate held-out dataset before the gate is used for new issues. Recommendations identify items that may be accepted under the policy and items that should remain under review.

How it works

The MCP process is started with a directory containing authorized issue JSON files. Jev is disabled by default. When enabled, the process can send selected title and body text to https://api.typesafe.ai/v1/systemone; expected labels are not sent. One request per issue asks three independent questions about the same issue, with no automatic retries.

Calibration creates a policy from a labeled report. Evaluation then checks that policy against a separate report. The implementation requires disjoint IDs between calibration and evaluation, and semantic duplicates should also be removed by the dataset owner. A policy can be disabled when the sample is small or the observed error is too high; disabled policies route every item to review without making model calls.

Setup and configuration

Install Python 3.11 or newer, install the package with its MCP extra, and start the server with an absolute data-root path, for example:

jev-gate-mcp --data-root /absolute/path/to/authorized-json

For Jev mode, provide TYPESAFE_API_KEY through the local environment or a secret manager. Start the process with --allow-jev and an optional --max-calls limit only when outbound model requests are authorized. The MCP bundle supports hosts that implement MCPB 0.4 with UV. A local demo can run without a key or network requests.

Tools and capabilities

Jev Decision Gate by BPJ provides six MCP tools covering the workflow described in the project documentation:

• Process local issue batches with the rules baseline or optional Jev provider.
• Check task and usage information before or during triage.
• Calibrate acceptance behavior from labeled examples.
• Evaluate a frozen policy on held-out data.
• Produce accept-or-review recommendations.
• Apply call limits and retain reports for later comparison.

The same implementation is also available through the Python package, command-line interface, and installable Skill. Reports can include provider failures and unknown usage rather than treating failed calls as free or successful.

Limitations and notes

The tool is an integration and statistical gate, not Jev itself, its training algorithm, or an official TypeSafe product. It does not mutate issue trackers: it never edits issues, applies labels, posts comments, or closes tickets. The fixed taxonomy is narrow and may require customization for repositories with different module structures.

The acceptance bound assumes independent and identically distributed samples. It does not account for distribution drift, adversarial inputs, or reviewer accuracy. The project makes no accuracy claim for Chinese or other languages; evaluation should use the language and issue types that matter to the repository.

TypeSafe bills Jev usage separately. Cost estimates based on reported provider input tokens are not invoices and exclude output, retries, review, fallback, and agent-host costs. The package has no telemetry, and the bundled examples are synthetic cases intended for trying the workflow.