KennyBot

A Flue-powered GitHub issue triage bot. Other repos consume it via a composite GitHub Action.

Quickstart

Add this workflow to your repo (.github/workflows/issue-triage.yml):

name: Issue Triage
on:
  issues:
    types: [opened, labeled]

jobs:
  analyze:
    if: |
      github.event.action == 'opened' ||
      (github.event.action == 'labeled' && github.event.label.name == 'auto-triage')
    runs-on: ubuntu-latest
    permissions:
      contents: read
      issues: read
      pull-requests: read
    outputs:
      decision: ${{ steps.triage.outputs.decision }}
    steps:
      - uses: actions/checkout@v4
      - id: triage
        uses: jpenilla/KennyBot/.github/actions/triage@main
        with:
          issue-number: ${{ github.event.issue.number }}
          model: opencode-go/deepseek-v4-flash
        env:
          OPENCODE_API_KEY: ${{ secrets.OPENCODE_API_KEY }}

  process:
    needs: analyze
    if: ${{ needs.analyze.outputs.decision != '' }}
    runs-on: ubuntu-latest
    permissions:
      issues: write
    steps:
      - uses: actions/checkout@v4
      - uses: jpenilla/KennyBot/.github/actions/process@main
        with:
          decision: ${{ needs.analyze.outputs.decision }}
          issue-number: ${{ github.event.issue.number }}
          trigger-label: auto-triage

Set the API key for your chosen model as a repo secret. See flueframework.com for available models and their env var names.

Open an issue — the bot will triage it.

Manual triage via label

Trigger triage on demand by adding a configured label to any issue. The process step removes the label automatically after the LLM finishes.

To enable, add labeled to the trigger types, add the if condition shown above, and set trigger-label to your chosen label name. The example above uses auto-triage.

Custom skill

Add a custom-skill-path input pointing at your own SKILL.md in your repo:

        with:
          issue-number: ${{ github.event.issue.number }}
          custom-skill-path: .github/kennybot-triage.md

Architecture

Two jobs, strict separation of concerns:

1. analyze (issues: read) — The Flue agent reads the issue, searches for duplicates via gh, and returns a structured decision.
2. process (issues: write) — A deterministic JS script parses the decision and calls Octokit. No LLM involvement.

Decisions

Decision	Action
leave-open	Comment (optional) + add labels, leave open
close-invalid	Comment explaining why, close
close-duplicate	Comment referencing the original issue, close
close-done	Comment noting already addressed, close