Jev Decision Gate · by BPJ

Use Jev for Issue triage without building the whole integration and review workflow.

中文 · MCP setup · Evaluation method · BPJ developer tools

Why install ours? Get six MCP tools for local Issue batches, task/usage checks, independent calibration, held-out evaluation and accept/review recommendations. The same workflow is available as a Skill or CLI. LLM vs direct Jev vs our MCP.

Underlying model prices: about 238× difference, with the same hypothetical input volume. Standard uncached prices on 2026-10-02: Claude Fable 5.1 $10/M input tokens; Jev 1.13 $0.042/M. Against Haiku 4.5 ($1/M), the ratio is 23.8×. Direct Jev and our MCP share this Jev input price. This compares model input unit prices; total workflow cost depends on actual usage, output, retries and agent-host charges.

Price math and assumptions · Static image · Actual rules output

An independent MIT-licensed Python tool for repository maintainers. Run a rules baseline, optionally call TypeSafe's Jev, and calibrate which suggestions may be accepted. The first case predicts issue kind, affected module and information sufficiency. It never edits issues, applies labels, posts comments or closes tickets.

Version 0.3.1, early developer preview. This is an integration and statistical gate, not Jev's model, training algorithm, or an official TypeSafe product. The bundled cases are 12 original synthetic examples for trying the workflow.

Run in two minutes

Python 3.11+:

git clone https://github.com/f-tiger/jev-decision-gate.git
cd jev-decision-gate
python -m venv .venv
# macOS / Linux
source .venv/bin/activate
# Windows PowerShell: .venv\Scripts\Activate.ps1
python -m pip install '.[mcp]'
jev-gate demo --out demo-report.json --export-issues demo-issues.json

The demo runs locally without keys or network requests. It produces real output from a deliberately simple rules baseline. All items initially require review. Keep the report: compare on your own labels before choosing a provider. Commands refuse to overwrite existing output files.

Reproduce the animation's price calculation without a key:

python tools/price_scenario.py --input-tokens 1000000 --fallback-fraction 0.1

This scenario is $0.042 for Jev input plus $1 for an extra 10% of that input sent to Fable, or $1.042 versus $10. The 9.6× input-cost ratio is hypothetical, excludes output/retries/review, and does not imply this package calls Fable. Share a successful run or a blocker.

Related MCP server: jev-eval-mcp

Call Jev on selected issues

Set TYPESAFE_API_KEY through your local environment or secret manager. Do not paste a key into an issue, model chat, repository or MCP configuration committed to Git. Jev mode sends selected title/body text to https://api.typesafe.ai/v1/systemone; BPJ receives nothing.

jev-gate triage demo-issues.json --provider jev --max-calls 12 --out jev-report.json

The default pins jev-1.13.0. One request per issue batches three independent questions against the same issue. There are no automatic retries. Failure returns a review recommendation and records unknown usage rather than assuming zero cost. Exit code 2 means an input failure or a saved report with provider failures; inspect stderr and the report.

To estimate the cost of provider-reported input tokens, explicitly add --input-usd-per-million YOUR_CURRENT_PRICE. Output is an estimate, not an invoice. Failed calls may be billed without reported usage. The estimate covers reported provider input; account for agent-host, review and fallback costs separately.

Bring your own cases

[
  {
    "id": "myrepo-123",
    "title": "Login fails after session expiry",
    "body": "Steps, environment, actual behavior and expected behavior...",
    "expected": {"kind": "bug", "module": "auth", "information": "sufficient"}
  }
]

expected is optional at inference time and required for calibration/evaluation. It is never sent to Jev. Use human-reviewed labels. Fixed labels in v0.3.1:

This taxonomy is intentionally narrow. Repositories whose modules do not fit should customize the question contract and recalibrate. No claim is made about accuracy in Chinese or other languages; evaluate the language you actually use.

Calibrate, then test separately

Run inference on independently labeled calibration and test files, then:

jev-gate calibrate calibration-report.json --out policy.json
jev-gate evaluate policy.json held-out-report.json --out evaluation.json
jev-gate triage new-issues.json --provider jev --policy policy.json --out decisions.json

The model, taxonomy and score mapping must be frozen before calibration. IDs must be disjoint across calibration and evaluation; semantic duplicates must also be removed by the dataset owner. Small samples or high error disable the gate. A disabled policy skips model calls and sends every item to review. Policy scope mismatches fail before inference.

The fixed threshold search uses an exact one-sided binomial bound with Bonferroni correction. The bound concerns errors among accepted predictions under i.i.d. sampling. It does not cover distribution drift, adversarial input or reviewer accuracy. Read the assumptions and score definition.

MCP, Skill and Python

Download the MCP bundle for hosts supporting MCPB 0.4 with UV. Choose your Issue JSON directory; Jev is off by default. Registered as io.github.f-tiger/jev-decision-gate in the official MCP Registry. Installation and distribution status.

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

Jev is disabled by default in MCP. Add --allow-jev --max-calls 20 at startup only when outbound model calls are authorized and the process can read your key. Six tools and client configuration.

Use the installable instructions under skills/jev-decision-gate/ in a compatible Skill host. Installation locations differ by host; this repository does not silently install or modify your host. The bundled Skill can run the same Python implementation without downloading project code during a task.

For hosts supported by the Skills CLI:

npx skills add f-tiger/jev-decision-gate --skill jev-decision-gate

Discovery, an isolated public-repository installation and the installed Skill’s rules demo have been verified. Choose your agent during installation. The third-party Skills CLI has its own optional telemetry; see its documentation for DISABLE_TELEMETRY=1. This package itself has no telemetry.

from jev_decision_gate.triage import triage
report = triage([{"id": "example-1", "title": "API error", "body": "Please investigate"}])

What is free? What is paid?

The package, MCP server, Skill, fixtures and evaluator are free under MIT. Jev usage is billed separately by TypeSafe according to your account. BPJ does not sell a hosted Decision Gate service in this preview. Recurring evaluation, policy history and team review are possible future paid features; demand and delivery have not been validated.

See BPJ developer tools for the existing website. No private source code or keys are required to browse it. The package has no telemetry; website visits and GitHub stars do not prove successful installations.

Verify and contribute

python -m unittest discover -s tests -v
python tests/smoke_mcp.py

Tests use mocks and synthetic fixtures, never paid APIs. Verification status, security boundaries, contributing and roadmap. Contributions are welcome for independently labeled, redistributable cases, failure reports and controlled comparisons with a rules baseline. Keep private issues and credentials out of public contributions.

Protocol reference: TypeSafe API, models, MCP Python SDK. Jev references describe interoperability; no affiliation is implied.

Want to share it? Use the launch copy and evidence links. AI-assisted implementation; no official TypeSafe affiliation.