Pr Jangler

PR Jangler

Autonomous PR backlog handler. Cron fires prj-orchestrator every N minutes; the orchestrator picks one queued action and dispatches to one phase skill. Per-PR state is durable; every transition is logged; nothing acts on a PR until a claim has been independently reproduced.

   DISCOVER       TRIAGE         VERIFY         REVIEW         PLAN            VALIDATE         IMPLEMENT      DECIDE         REPORT
  ┌────────┐    ┌────────┐    ┌────────┐    ┌────────┐    ┌──────────┐    ┌──────────────┐   ┌──────────┐   ┌────────┐    ┌────────┐
  │ GH → Q │───▶│Classify│───▶│Reproduce│──▶│ Read   │───▶│Fix +     │───▶│Adversarial   │──▶│ New PR   │──▶│ Label  │───▶│ Email  │
  │ sync   │    │ PR/cmt │    │ claim   │   │ diff   │    │test plan │    │6-item check  │   │ to branch│   │ verdict│    │ digest │
  └────────┘    └────────┘    └────────┘    └────────┘    └──────────┘    └──────────────┘   └──────────┘   └────────┘    └────────┘
  prj-discover  prj-triage    prj-verify    prj-review    prj-plan-fix    prj-validate-      prj-implement   prj-decision   prj-report-
                              -claim                                       adversarial        -fix                          daily

  Also dispatched on cadence: prj-detect-overlap (duplicate scoring), prj-setup (one-shot install).
                              Orchestrated by: prj-orchestrator (cron heartbeat).

Operating Principles

These are non-negotiable across every phase skill. Hoisted here so each child does not redefine them.

1. The queue is authoritative. {project-root}/_bmad-output/pr-workflow/state.json is the single source of truth. Phase skills NEVER read GitHub for queue facts — they read state.json. Only prj-discover writes from GitHub into the queue.
2. All state I/O goes through state_io.py. Atomic via tmp + fsync + rename. Direct file writes to state.json are forbidden — a corrupted queue blocks the whole pipeline.
3. Idempotent per heartbeat. Every action survives mid-run interruption: the next heartbeat picks up cleanly without double-acting or losing the transition.
4. Never silently fail. Every outcome — dispatch, stub, abort, error, no-op — emits a structured JSONL entry to {project-root}/_bmad-output/pr-workflow/logs/{YYYY-MM-DD}.jsonl. Errors also surface on stderr.
5. PleaseAdvise is the escape hatch. Any phase may transition a PR to PleaseAdvise (priority 1000, outranks everything). Stays there until the user-acknowledged override label is applied.
6. Never push to the contributor's branch. prj-implement-fix opens a new PR targeting the contributor's branch. Force-pushes are forbidden. Every fix is a new commit; every fix is a new PR.
7. Verify before acting on any claim. A comment asserting a bug is a hypothesis, not a fact. prj-verify-claim reproduces in an isolated worktree before prj-plan-fix is allowed to run.
8. System actions outrank per-PR actions. Daily report (priority 700) and discovery sweep (priority 500) preempt the per-PR backlog (base 100).

Routing Table

When the orchestrator's next action — or a user invocation — names one of these signals, dispatch to the matching child skill.

Common Combinations

The canonical per-PR pipeline, in order:

prj-discover → prj-triage → prj-verify-claim → prj-review → prj-plan-fix
            → prj-validate-adversarial → prj-implement-fix → prj-decision

Not every PR walks the full sequence. Triage may shortcut: a duplicate goes straight to prj-detect-overlap → prj-decision (close-as-not-now); a clean review with no claim goes prj-review → prj-decision (ready-to-merge); a verification failure goes prj-verify-claim → prj-decision (request-changes-with-pushback).

Cross-Cutting Conventions

State tree layout

Owned by prj-orchestrator. Every phase reads from and writes to it (via state_io.py).

{project-root}/_bmad-output/pr-workflow/
├── state.json                       # the queue
├── prs/{pr_number}/                 # per-PR cache
│   ├── meta.json
│   ├── triage.md
│   ├── verification.md
│   ├── review.md
│   ├── overlap.md
│   ├── fix-plan.md
│   ├── adversarial.md
│   └── decisions.log                # JSONL audit
├── logs/{YYYY-MM-DD}.jsonl          # run-log
├── reports/{YYYY-MM-DD}.md          # daily-report archive
└── worktrees/                       # ephemeral, owned by prj-verify-claim

Configuration

Module config lives under [modules.prj] in {project-root}/_bmad/config.toml (or config.yaml per the active BMad version). Required: prj_repo. Common: prj_discover_every_n, prj_report_hour_local, prj_smtp_host, prj_smtp_port, prj_email_to, prj_bot_user. User-only keys (user_name, communication_language) live in config.user.yaml.

Path tokens

{project-root}/... is a literal token in config values. Do not substitute it. Filesystem operations resolve it at the point of use; the config itself keeps it symbolic.

Dependency map

Label vocabulary

Pre-created on prj_repo by prj-setup (from skills/prj-setup/assets/labels.txt). Every phase that writes to GitHub uses these labels — never invents new ones.

• prj/phase:{Discovered,Triaged,VerifyClaim,Reviewing,FixPlan,AdversarialCheck,FixImpl,Decision,ReadyToMerge,Blocked,Rejected,Archived,PleaseAdvise}
• prj/decision:{ready-to-merge,request-changes,close-as-not-now}
• prj/duplicate-of:{n}, prj/please-advise, prj/needs-retriage, prj/user-ack

Install Modes

This skill is the hub. There are three ways to use PR Jangler:

1. Full pack — install the whole plugin. The hub plus all 12 phase skills land in .claude/skills/. Recommended for autonomous operation.
2. Hub only — install just pr-jangler/. You get the routing knowledge, conventions, label vocab, and operating principles, but no executable scripts. Useful for orienting an agent that will hand off to a separately-installed phase skill, or for understanding the workflow before committing all 12.
3. À la carte — copy individual prj-* skills into your project. Skip the hub; each child SKILL.md is self-contained for its phase.

When this hub is loaded and a user task names a phase action that requires an uninstalled child, point them at the missing skill rather than improvising it: "this needs prj-verify-claim, which is not installed at .claude/skills/prj-verify-claim/. Install via <plugin install line> or copy from the pack repo."

Out of Scope

This hub routes the PR Jangler module. It does NOT:

• Review PRs in repos other than prj_repo. The configured repo is the workspace; multi-repo orchestration is not built.
• Write or modify skills. Skill authoring belongs to skill-creator, not here.
• Handle non-PR GitHub events (Issues, Discussions, Actions failures). Use a dedicated issue-triage skill if you need that.
• Replace BMad core or other modules. It runs alongside them under _bmad/.
• Generate code changes from scratch. It can plan and implement fixes for verified bugs on existing PRs — not greenfield feature work. Use bmad-bmm-dev-story or incremental-implementation for that.
• Bypass the verification gate. A claim that cannot be reproduced gets a polite pushback comment, not a speculative fix.