Why it exists

Existing LLM triage bots are too slow and too expensive to run on every issue, and they hand back prose you still have to parse. Jev is a System One model: text in, typed probabilistic decisions out - choice, score, noul - each with a calibrated confidence, in roughly 70-500 ms at near-zero cost.

What it does

On every opened or edited issue, jev-triage asks a small set of bounded questions (type, priority, area, is-it-security) and applies labels when confidence clears the threshold. Uncertain cases get triage:needs-human instead of a guess. It’s the “System One in front of System Two” cascade: cheap and fast on everything, honest about what it doesn’t know.

Highlights

Typed, confidence-gated labels - never a silent wrong guess
A security-flag path that alerts rather than labels
Backlog mode to triage existing issues in bulk
Gated by Foundry, like every tool in the family

jev-triage

Fast, near-free GitHub issue triage - powered by TypeSafe AI's Jev.

Jev is a System One model: text state in, typed probabilistic decisions out (choice / score / noul), each with a calibrated confidence, in ~70-500 ms at roughly free cost. That combination makes it viable to run on every issue the moment it opens - and to escalate only the ones it's unsure about to a human or a real LLM.

# .github/workflows/triage.yml
on:
  issues: { types: [opened, edited, reopened] }
permissions: { issues: write, contents: read }
jobs:
  triage:
    runs-on: ubuntu-latest
    steps:
      - uses: cmaintz/jev-triage@v0
        with:
          jev-api-key: ${{ secrets.JEV_API_KEY }}
          cloudflare-account-id: ${{ secrets.CLOUDFLARE_ACCOUNT_ID }}

--- run summary ---
#142  label=bug, priority:major   conf 0.91   ✓ applied
#143  type uncertain              conf 0.48   → triage:needs-human
#144  security flagged            noul 0.83   ⚠ alerted
1,204 issues triaged this month · $0.11

Why this exists

Existing LLM triage bots are too slow and too expensive to run on every issue, and they hand you prose you still have to parse. Jev returns a typed value your workflow branches on and a confidence you gate on. The result is triage that's cheap enough to always run and honest enough to defer when it isn't sure.

How it works

On a new/edited issue, the title + body become the Jev state.
Your .github/jev-triage.yml defines the questions - one batched call answers them all (adding questions is near-free).
Confident answers become labels / routing; low-confidence answers are never applied - they get triage:needs-human or escalate to an LLM.
A run summary reports labels, escalations, tokens and estimated cost.

See examples/jev-triage.yml for the full config surface.

Beyond per-issue triage

Backlog sweep - trigger on workflow_dispatch / schedule and it map-reduces every open issue (skipping PRs and anything already carrying the marker_label), capped by max_issues. Sweep a whole backlog for pennies. See examples/backlog-workflow.yml.
LLM escalation cascade - with on_low_confidence: escalate and an llm-api-key, low-confidence issues are re-classified by an LLM (any OpenAI-compatible endpoint via llm-model / llm-base-url), which then labels and leaves a one-line rationale. Jev is System One; the LLM is System Two - invoked only where Jev is unsure.
Duplicate detection - set dedupe: true: GitHub search retrieves candidate issues, and Jev picks the duplicate (or "none") from that bounded set. A match gets a possible-duplicate label + link (it never auto-closes). Retrieval is code; judgment is Jev.

Honest limitations

A first-pass, not a decision-maker. Jev is ~68% accurate on classification - fast and cheap, not smarter. The confidence gate is the point. It never auto-closes issues.
Text only. Screenshots and attachments carry no signal to Jev.
No arithmetic / dates. Those stay in code (e.g. routing is a deterministic map, not a Jev question).
Needs a Jev key. Sign up at TypeSafe for a first-party key, or use Cloudflare Workers AI (typesafe/jev). First-party is the default (one secret).

Status

v1.0 - per-issue triage, backlog sweep, LLM escalation cascade, and duplicate detection. Passes the Foundry gate (mise run gate: lint → typecheck → test → audit), 28 unit tests, bundles clean. Verified against the official API reference and Cloudflare's model page, and validated against the real Jev API (a live test classifies a real bug report end-to-end; test/live.test.ts). Contributions welcome. See CHANGELOG.md and the full spec.

Development

Quality is enforced through Foundry's six-verb gate:

mise run gate   # lint → typecheck → test (coverage floor) → audit - the oracle
mise run fix    # auto-fix (eslint --fix + prune suppressions)
npm run build   # bundle dist/index.js (ncc) - commit the result