Backporcher

Backporcher


A fully autonomous software engineering pipeline. Label a GitHub issue with `backporcher`, and in ~20 minutes you get a merged PR with tests passing and the issue closed. A real-time web dashboard lets you manage the fleet: approve or hold tasks before merge, pause/resume the dispatch queue, re-run failed agents, and monitor every stage of the pipeline from triage to merge.

Built in early 2026. 100% auto-merge rate on its first production run (15 PRs, zero manual interventions). This mirrors the agent orchestration architectures emerging from Anthropic's Claude Code and Augment's multi-agent systems, but as a standalone, open-source daemon you can run on your own infra.

What makes it different

Most "AI coding" tools are glorified autocomplete. Backporcher is an **end-to-end pipeline**: it triages, plans dependencies, dispatches sandboxed agents, reviews their work with a coordinator agent, retries CI failures with error context, and merges, all autonomously.

The key insight: treat agents like junior developers. Give them isolated worktrees, review their PRs, and let CI be the final gate. But unlike junior developers, give them a **code-aware navigation map** so they don't waste time exploring the codebase, and feed them **learnings from every past success and failure** in that repo so they get better over time. No magic, just good engineering around `claude -p`.

The Pipeline

GitHub Issue (label: backporcher)
  → Haiku triages complexity (sonnet vs opus)
    → Batch orchestrator assigns priorities + dependency chains
      → Sonnet queries code graph → navigation map of relevant files/symbols
        → Sandboxed claude -p in git worktree (with stack info + learnings + navigation map)
          → Build verification (optional, per-repo)
            → PR created
              → Code graph builds blast radius (Tree-sitter + dependency BFS)
                → Coordinator reviews diff + impacted c