pr-jangler

PR Jangler

An autonomous workflow for your GitHub PR backlog.

Cron fires a heartbeat. The heartbeat picks one queued action. That action dispatches to one phase skill. Repeat forever. The backlog clears itself — triaged, verified, reviewed, sometimes fixed (via a new PR targeting the contributor's branch), always logged, never silently failing.

   DISCOVER     TRIAGE       VERIFY       REVIEW       PLAN          VALIDATE       IMPLEMENT    DECIDE       REPORT
  ┌────────┐  ┌────────┐  ┌─────────┐  ┌────────┐  ┌──────────┐  ┌────────────┐  ┌──────────┐  ┌────────┐  ┌────────┐
  │ GH→Q   │─▶│Classify│─▶│Reproduce│─▶│Read    │─▶│Fix +     │─▶│Adversarial │─▶│ Fix-PR   │─▶│ Verdict│─▶│ Email  │
  │ sync   │  │ PR/cmt │  │ claim   │  │diff    │  │test plan │  │6-item gate │  │ to branch│  │ + label│  │ digest │
  └────────┘  └────────┘  └─────────┘  └────────┘  └──────────┘  └────────────┘  └──────────┘  └────────┘  └────────┘

---

Three install modes

PR Jangler ships as three artifacts so you can pick the dose that fits.

Mode	What lands	Best for
Full pack	Hub + 12 phase skills	Autonomous operation under cron
Hub only	pr-jangler skill (routing knowledge, conventions, principles) — no scripts	Orienting an agent, lightweight understanding
À la carte	Pick individual prj-* skills	One-off use of a specific phase

Full pack — Claude Code

/plugin marketplace add delorenj/pr-jangler
/plugin install pr-jangler@pr-jangler-marketplace

Hub only — Claude Code

/plugin install pr-jangler-hub@pr-jangler-marketplace

The hub teaches an agent the workflow, the state-file layout, the label vocabulary, and which phase to call for which signal. It will tell you which child skill to install if a task needs executable scripts.

Local development

git clone https://github.com/delorenj/pr-jangler.git
claude --plugin-dir /path/to/pr-jangler

À la carte (any agent)

Every skills/prj-*/SKILL.md is self-contained for its phase. Copy the directory into your agent's skills path (.claude/skills/, .cursor/rules/, .kiro/skills/, etc.) and it works without the hub.

---

The 12 phase skills

Skill	Phase	One-liner
prj-setup	Install	Writes config, creates GH labels, validates gh/op/SMTP, initializes the workflow tree
prj-orchestrator	Heartbeat	Cron entrypoint; picks one action, dispatches one phase, persists state
prj-discover	Discover	Syncs GitHub open-PR set into the queue; adds, archives, bumps comment counters
prj-triage	Classify	Dual-mode: classifies a PR (--mode pr) or the latest unclassified comment (--mode comment)
prj-verify-claim	Verify	Reproduces a comment's alleged bug in an isolated worktree, or pushes back politely
prj-review	Review	Reads diff + changed files; produces review.md with structured findings
prj-detect-overlap	Overlap	Scores file-footprint overlap vs every open PR; LLM-compares the high-overlap pairs
prj-plan-fix	Plan	Designs a failing test + smallest diff + rationale + risks → fix-plan.md
prj-validate-adversarial	Validate	Runs full regression suite + 6-item adversarial checklist; default verdict is reject
prj-implement-fix	Implement	Opens a new PR targeting the contributor's branch. Never push to their branch. Never force-push.
prj-decision	Decide	Aggregates the per-PR cache, applies the rubric, labels and comments the verdict
prj-report-daily	Report	Renders an HTML+text email digest, sends via SMTP, archives to reports/

The hub (pr-jangler) sits on top of these and owns the cross-cutting rules nobody wants to repeat 12 times: state-file layout, label vocabulary, idempotency rules, dependency map, the PleaseAdvise escape hatch.

---

Architecture in one paragraph