JEV Model Navigator

Route Issues and Pull Requests to the right AI model using TypeSafe Jev as a typed decision layer inside GitHub Actions.

Teams often hard-code a single model for every task. This Action collects Issue/PR context (and optional compact PR file metadata), asks Jev to choose from an allowlisted catalog, then returns stable outputs your workflow can branch on—without invoking the selected model by default.

- id: nav
  uses: JevForge/jev-model-navigator@v0.4.1
  env:
    AI_GATEWAY_API_KEY: ${{ secrets.AI_GATEWAY_API_KEY }}
  with:
    model_catalog_path: .jev/model-catalog.yml

Note: Output provider is the selected model vendor (for example openai). It is not jev_provider (how this Action reaches Jev).

• Typed model routing powered by Jev (experimental_evaluate, not free-form generation)
• Secret-based authentication (AI_GATEWAY_API_KEY, TYPESAFE_API_KEY, or JEV_CUSTOM_API_KEY)
• Structured outputs for later workflow steps (selected_model, decision, confidence, …)
• Optional PR diff metadata (paths and line counts only—never patch hunks)
• Optional Issue/PR comments, managed jev:* labels, and Check Runs
• Deterministic allowlist + schema validation; Jev prose is never executed
• Configurable low-confidence policy: fail | warn | request-review | no-op

Issue / Pull Request
        ↓
Collect title, body, optional PR file metadata
        ↓
Sanitize + derive task signals
        ↓
Jev evaluates allowlisted candidates
        ↓
Schema + confidence policy
        ↓
Action outputs (+ optional comment / labels / check)
        ↓
Next CI/CD step

flowchart LR
  A[GitHub Event] --> B[Collectors]
  B --> C[Jev]
  C --> D[Validate]
  D --> E[Policy]
  E --> F[Outputs]

1. Read task text from the Issue/PR (or task input).
2. Load candidates from candidates, .jev/config.yml, or model_catalog_path.
3. Call Jev through jev_provider (no silent provider fallback).
4. Validate against an allowlist and schema; optionally ask for alternate_model.
5. Apply low_confidence_policy.
6. Emit outputs. Free-form explanation text is display-only.

Pull Request: "Refactor auth middleware"
        ↓
Diff signals: TypeScript + auth/ paths, +120/-40
        ↓
Jev → selected_model = openai-gpt-5-5
      alternate_model = google-gemini-3-5-flash
      decision = SELECT_MODEL
      confidence = 0.86
        ↓
Downstream job routes the coding agent to GPT-5.5

1. Copy examples/model-catalog.yml to .jev/model-catalog.yml in your consumer repo.
2. Add repository secret AI_GATEWAY_API_KEY (default provider).
3. Add a workflow:

name: Model Navigator
on:
  issues:
    types: [opened, edited]
  pull_request:
    types: [opened, edited, synchronize]

permissions:
  contents: read
  issues: write
  pull-requests: write
  checks: write

jobs:
  navigate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - id: nav
        uses: JevForge/jev-model-navigator@v0.4.1
        env:
          AI_GATEWAY_API_KEY: ${{ secrets.AI_GATEWAY_API_KEY }}
        with:
          model_catalog_path: .jev/model-catalog.yml
          decision_only: 'true'
          include_pr_diff: 'true'
          create_check_run: 'true'

      - name: Use recommendation
        run: |
          echo "decision=${{ steps.nav.outputs.decision }}"
          echo "model=${{ steps.nav.outputs.selected_model }}"
          echo "fallback=${{ steps.nav.outputs.alternate_model }}"

Pin @v0.4.1 for reproducibility, or @v0 for the floating major line.

Marketplace listing: JEV Model Navigator

See examples/workflows/model-navigator.yml and:

• examples/basic.yml — minimal outputs-only run
• examples/pr.yml — PR with diff signals + Check Run
• examples/gate.yml — branch later steps on decision

- name: Continue with selected model
  if: steps.nav.outputs.decision == 'SELECT_MODEL'
  run: echo "Use ${{ steps.nav.outputs.selected_model }}"

- name: Request human review
  if: steps.nav.outputs.decision == 'REQUEST_REVIEW'
  run: echo "Navigator asked for review"

- name: Fail closed on abstain
  if: steps.nav.outputs.decision == 'ABSTAIN'
  run: |
    echo "No safe candidate"
    exit 1

Create a repository secret:

Repository → Settings → Secrets and variables → Actions → New repository secret

Never put API keys in workflow YAML, logs, or Issues. There is no silent fallback between providers.

Optional local config: .jev/config.yml (inputs override file values).

Jev is TypeSafe’s evaluation model for structured decisions, not chat. This Action needs a choice from an allowlist of model ids, confidence, and optional abstain/review signals. Jev returns typed answers (choice / boolean) that code can validate. Generative text would be unsafe to treat as a command or model id. That is why the Action uses experimental_evaluate and never generateText for the decision.

Only:

• sanitized Issue/PR task text (truncated; secrets redacted)
• derived task signals
• optional PR diff metadata (paths, languages, +/- counts, sensitive/test/infra flags)
• candidate metadata (id, provider, capabilities, context, cost tier, availability)
• budget preference and confidence constraints

Never: GitHub tokens, API keys, patch hunks, or unrelated checkout blobs.

Outputs only:

permissions:
  contents: read

Comments, labels, Check Runs, and PR file listing:

permissions:
  contents: read
  issues: write
  pull-requests: write
  checks: write

uses: JevForge/jev-model-navigator@v0.4.1   # recommended pin
uses: JevForge/jev-model-navigator@v0       # floating major (v0.x)

You do not need to create the GitHub Release by hand:

1. Merge to main with dist/ up to date (npm run build).
2. Either:
   • Actions → Release → Run workflow on main, enter 0.4.1, or
   • git tag v0.4.1 && git push origin v0.4.1
3. CI runs tests, publishes the GitHub Release, and moves the floating major tag (v0).

Marketplace listing updates still need one browser step (GitHub 2FA): open the release and keep Publish this Action to the GitHub Marketplace checked.

See CHANGELOG.md and Releases.

Requires Node.js 24+.

npm install
npm test
npm run typecheck
npm run build

Consumers run bundled dist/index.js (runs.using: node24) and do not need to install dependencies.

Optional local planner CLI (fixtures in this repo):

npm run batch:plan

See CONTRIBUTING.md. Report bugs via Issues—never include API keys or tokens.

See SECURITY.md.

MIT