Why

Maintaining an active repository is endless manual work: triaging issues, spotting duplicates, checking what was already fixed, reviewing PRs, watching CI, resolving conflicts, keeping docs current. RepoHeart automates the repetitive parts while keeping humans in control of anything risky.

It runs entirely inside your GitHub Actions workflow — no hosted server, no database, no external infrastructure. Add one workflow, choose an AI provider, done.

How It Works

GitHub Event → Actions runner → normalize + dedup → route to agents
            → each agent retrieves bounded context, reasons, proposes actions
            → Safety Gate authorizes every write
            → results written back to GitHub → run exits (holds no state)

Deterministic where it must be (routing, permissions, idempotency), AI-driven where it helps (the reasoning inside each agent).

Quick Start

1. Add the workflow — .github/workflows/repoheart.yml:

name: RepoHeart
on:
  issues:
    types: [opened, edited]
  pull_request:
    types: [opened, synchronize, reopened]
  workflow_run:
    types: [completed]

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

jobs:
  repoheart:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: OpenAgentHQ/repoheart@main
        with:
          config: repoheart.yml

2. Configure — repoheart.yml:

repoheart:
  provider:
    name: opencode          # opencode | claude | openai | gemini | local
    model: your-model
  agents:
    issue_triage: true
    duplicate_detection: true
    pr_review: true
    code_quality: true
    security: true
    ci_repair: true
    conflict_resolution: true

3. Add provider credentials as repository secrets (e.g. OPENCODE_API_KEY, ANTHROPIC_API_KEY, OPENAI_API_KEY).

That's it.

The Agents

Enable only what you want in repoheart.yml.

Switching AI Providers

Change one line. Agents are untouched.

provider:
  name: claude      # was: opencode
  model: claude-model

Built for Big Repos Too

RepoHeart doesn't load your whole codebase into a prompt. It retrieves only what each event needs — using tree-sitter, ripgrep, and GitHub Search to do the heavy lifting, with event-scoped (sparse/shallow) checkout and an optional content-hash cache. Cost scales with the event's blast radius, not your repo size.

Safety First

Safety first, then correctness, then automation.

• Every write passes a mandatory Safety Gate.
• Agents can never escalate their own permissions.
• No force-push. No auto-merge in the MVP. No committed secrets.
• Low-confidence or high-risk situations escalate to a human with a clear explanation — never a silent guess, never broken state left behind.

Documentation

Status

v1.0 — Fully Implemented. All 7 roadmap phases complete: deterministic core, provider abstraction, issue intelligence, PR intelligence, large-repo scaling, CI repair & conflict resolution, and documentation agent. 524+ tests passing, ruff clean, mypy strict clean, provider switching config-only. One-workflow + one-config onboarding verified on real repos. See ROADMAP.md for full details.

License

See LICENSE.