Repo-agnostic dev-loop + tooling skills for Claude Code — stacked-PR ops, a bounded iteration-loop engine, and a diagnostic toolkit. It is forge's single companion plugin: forge resolves every agent capability it consumes (the iteration loop, the PR ops, the diagnostics) to a skill here. Each skill is also useful standalone, with no forge involved.
It infers PR topology from GitHub (gh) by default and falls back to raw git
analysis (tracked upstreams, merge-bases) when gh is unavailable. No
dependency on other plugins.
| Skill | Purpose |
|---|---|
/restack |
Bring a PR's base into its branch — sync the base with upstream first, then merge (default) or rebase (--rebase). Stops on conflict; never force-pushes a base. |
/restack-all |
/restack across an entire stack, bottom-up — each PR onto its just-updated base. Stops at the first broken layer. |
/deslop |
Strip AI slop from a PR diff — delete code-restating comments, cap survivors at one line, collapse over-complex local code (reusing /simplify). --protect <globs> spares designated files. |
/pr-brief |
Write/refresh a tight 1–3 sentence PR description, idempotently spliced into a marker-bounded body region. |
| Skill | Purpose |
|---|---|
/review |
Multi-channel PR review — parallel lens fan-out (lens-reviewer agent + lens pool) + always-on /code-review and /security-review → one verdict. |
/review-watch |
Persistent PR monitor firing a --on-trigger handler on each new review-like event; re-arms until stopped. |
/address-review |
Drive externally-submitted reviewer feedback to resolution — intake, triage, fix-walk under --protect, reply/resolve, re-request. |
/request-review |
Rank the most relevant peer reviewer by signal precedence and, gated, mark ready + request. |
/author-review |
Guide the author's self-review — structured diff walkthrough + manual-verification pass (via a repo how-to), each embeddable as an idempotent collapsible body section. |
/ci-green |
Drive a GitHub PR's CI to green — bounded fix-to-green loop, optional --until-merge monitor. |
/find-blocker |
Classify whether a PR is held by an external blocker (red/behind base, infra, sibling PR); emit a neutral, waitable condition spec. |
| Skill | Purpose |
|---|---|
/grind |
General bounded-autonomy iteration loop toward any verifiable target — grind to green, commit each step, stop at success/budget. Designed to be wrapped by more specific fix-loops. |
/stuck-check |
In-loop rabbit-hole detector — score loop signals, return continue / raise-threshold / halt. |
/root-cause |
Hypothesis-driven root-cause analysis with parallel fan-out. |
/hypothesize |
Lighter 2–4 candidate hypothesis loop with one cheap experiment per round. |
/pepper |
Scatter uniquely-marked trace logs through suspect code, run the repro, grep, iterate, clean up. |
/trace |
Route verbose process output to disk and grep it instead of polluting agent context. |
/restack # restack the current branch's PR on its base
/restack-all # restack the whole stack
/deslop # de-slop the current PR's diff
/review # multi-channel review of the current PR
/grind "<verifiable target>" # grind any target to green
/root-cause # find the root cause of a bug, flake, or regression
/restack honors your merge-vs-rebase preference (persona / git discipline);
with none stated it merges. Requires git and, for topology inference, the gh
CLI.
MIT.