agentic-devkit

Review one or more GitHub PRs in an isolated git worktree using parallel agents, without disturbing the working repo. Local routine: for each PR it creates a dedicated review worktree on a review-pr-<N> branch (never your own working branch), runs the project's PR-review skill, optionally auto-posts the comments that clear a high bar, then tears the review worktree down. Use when asked to review a PR, review my assigned PRs, do a worktree PR review, or when a scheduled "PR review" routine fires. Parameterized: pass repo (full local clone path for unattended runs), pr (a number, a list, or "mine" = all PRs assigned to me), and optionally the review skill, parallelism, cleanup, and post (auto = post the comments that clear a high bar without asking, THE DEFAULT; draft = review only, post nothing). Invoke as `run a_r_l_pr_review for repo=<path> pr=<number|mine>`. Triggers even without the exact name: "review PR 123 in a worktree", "review the PRs assigned to me", "do an isolated review of this PR".

Review a GitHub PR end-to-end from just its URL. Give it a PR link (or owner/repo#N); it finds the repo you ALREADY have cloned locally (via a cached lookup, then a scan of your cd_w workspace — never a duplicate clone), clones into cd_w only if you don't have it, spins up a git worktree checked out on the PR's real head branch updated to latest, then runs the project's own review skill (review-pr) if it has one or the global global-pr-reviewer otherwise, auto-posts the comments that clear a high bar as GitHub inline review comments, and tears the worktree + local branch down. This is a SKILL, not a routine (no _r); a routine may call it. Use when asked to "review this PR <url>", "review a PR from its link", "do a full review of <github pull url>", or given a bare GitHub PR URL to review. Parameterized: pr (URL / owner/repo#N / number), post (auto | draft), reviewer (auto | project | global).

Unattended QA routine that smoke-tests a web flow on STAGING and files ONLY confirmed, reproducible bugs to a Jira epic. Local routine: it reuses the signed-in browser session, drives a named flow skill end-to-end on desktop and mobile viewports, applies a strict settle-and-reproduce verification gate, dedupes against the epic's existing children, and files clean bug tickets. Use when asked to smoke-test / QA a staging flow, hunt for reproducible bugs in a booking or signup funnel, or when a scheduled "staging QA" routine fires. Parameterized: pass flow_skill, base_url, epic, and ticket_creator. Invoke as `run a_r_l_staging_qa_sweep with flow_skill=<skill> base_url=<staging host> epic=<KEY>`. Triggers even without the exact name: "QA the app flow on staging and file bugs", "smoke test the booking funnel", "find reproducible defects on staging".

Create a clean, human-readable git commit for the current changes. Groups the diff into a coherent commit, writes a clear message in the project's style, and commits (respecting the repo's permission posture). Say "a_sk_commit" or "commit" when ready to commit. Generic and project-agnostic — imported from an upstream framework and de-coupled from it.

Open a GitHub pull request for the current branch, filling the project's PR template properly. Pushes the branch if needed, fills the template body from the actual changes, picks the correct base branch, and creates the PR (respecting the repo's permission posture). Say "a_sk_pr" or "pr" when ready to open a PR. Generic and project-agnostic — imported from an upstream framework and de-coupled from it.

Drive test coverage on the current change up to the project's SonarQube (or CI) coverage gate. Finds the coverage command, measures coverage on the changed/new code, writes tests following the project's testing conventions for the uncovered lines, and re-runs until the gate passes. Say "a_sk_sonarqube_coverage", "raise coverage to the gate", or "get coverage green". Generic and project-agnostic — imported from an upstream framework and de-coupled from it.

Write professional work messages (Slack, email, escalations, status updates) from the standpoint of a VP of Engineering / VP R&D. Use when the user asks to draft, rewrite, sharpen, or "make this sound better" for a message to colleagues, managers, clients, or cross-team threads, especially under deadline pressure or conflict. Protects delivery teams, makes trade-offs visible, stays factual and non-defensive.

Turn a rough, half-formed task description into a clean, self-contained instruction prompt for an autonomous or scheduled routine (Claude Code local/cloud routine, /goal, /loop, /schedule, or a reusable skill). Vets the raw text, fixes typos and ambiguity, clarifies only the decisions that change the steps, verifies the plan against a prior real run when a reference exists (a PR, a past routine, a transcript), and hardens the result for unattended execution. Use when the user says "write this routine instruction", "make this prompt count", "clean up this prompt so Claude does it right", pastes a draft for a scheduled/local routine, or asks to author instructions for a recurring/automated task.