Skip to content

Latest commit

 

History

History
85 lines (68 loc) · 8 KB

File metadata and controls

85 lines (68 loc) · 8 KB

toolkit

A direct-use Claude Code plugin from Agent Workshop: six curated agents (code review, governance, and CI monitoring) plus eleven direct-use skills you can run in any repo with no setup. The agents read your code, specs, tests, and CI and report findings — they never modify your files. Three skills produce structured handoff artifacts (review briefs, PR opens, and goal documents for a new session to pursue); doc-to-html renders a markdown report as a standalone dark HTML page; claim-check runs an unbiased, evidence-grounded investigation of a premise and returns a verdict plus a readiness dossier; qa-sweep fans a QA team over a broad surface and corroborates every finding firsthand before it counts; empirical-proof proves a just-finished change at the running software — real MCP/HTTP calls, probe scenarios, raw evidence — and reports verified / broken / blocked without fixing anything; code-quality-review runs an unusually strict, structure-first maintainability review over a branch's diff and pushes for restructurings that delete complexity rather than rearrange it; get-pr-comments fetches and triages the active PR's review comments into a prioritized action list (read-only — it never replies unless you ask); ui-demo-video records a Playwright walkthrough of the running app after UI work — per-scene PNG frames the model reads to verify the UI, plus a shareable mp4 for the PR; route-work grades a task about to be dispatched and returns a model + effort + process-pattern route, carrying the canonical model × effort table as its single source of truth.

Install

In a Claude Code session, add this repo as a marketplace, then install the plugin:

/plugin marketplace add giostriquer/agent-workshop
/plugin install toolkit@agent-workshop

(Terminal equivalent for the first step: claude plugin marketplace add giostriquer/agent-workshop.)

For Codex, use the skill-based counterpart:

codex plugin marketplace add giostriquer/agent-workshop --ref main
codex plugin add toolkit@agent-workshop

Codex plugins do not currently expose standalone custom agents from plugin manifests. The Codex toolkit package exposes handoff-review, handoff-pr, handoff-goal, doc-to-html, claim-check, qa-sweep, empirical-proof, code-quality-review, get-pr-comments, ui-demo-video, and route-work as skills and bundles the agent files inertly; use the agent-workshop onboarding plugin when you want to copy true .codex/agents/ wrappers into a target repo.

For Cursor, import this repo as a Team Marketplace (Teams/Enterprise, admin) — Dashboard → Settings → Plugins → Team Marketplaces → Add Marketplace → Import from Repo (giostriquer/agent-workshop) — then install toolkit from Customize in the sidebar.

After install, the six agents are available, namespaced toolkit:<agent> — e.g. toolkit:spec-reviewer. The eleven skills are available as handoff-review, handoff-pr, handoff-goal, doc-to-html, claim-check, qa-sweep, empirical-proof, code-quality-review, get-pr-comments, ui-demo-video, and route-work (skills are invoked by name, not namespaced). The same marketplace also hosts the agent-workshop onboarding plugin (/plugin install agent-workshop@agent-workshop) for the full scaffold-adoption flow.

Agents

Agent Inspects
spec-reviewer a design spec or implementation plan for gaps, before you build
code-quality-reviewer a code diff for maintainability and structure — the 1k-line rule, spaghetti growth, code-judo simplification; loads the code-quality-review skill's rubric and runs before pattern-reviewer
test-quality-reviewer a test diff (or existing tests) for trustworthiness and risk coverage
pattern-reviewer a code diff for implementation-pattern conformance; with no documented conventions it infers patterns from sibling files and labels findings lower-confidence
vigil a repo's agent / skill / workflow layer for governance drift
ci-watcher the current branch's PR CI — watches the checks and reports pass/fail with the failing-log excerpt or check link; read-only, background-friendly

All six are advisory and read-only (no Edit/Write) — the reviewers use Read, Grep, Glob, Bash; ci-watcher uses Bash, Read.

Skills

Skill Produces
handoff-review a self-contained brief for a separate agent/session to independently verify a branch (task-vs-code, rules, info-leak, correctness) before a PR — and, in continue mode, continue the work from a verified foundation
handoff-pr a structured PR handoff artifact (title, body, ticket links, status) for a separately-authorized session to open — never opens the PR itself
handoff-goal a self-contained goal contract directory — goal.md (frozen: outcome, baseline, acceptance checks with a real-surface primary verifier, integrity rules, approval gates, operating rules) plus plan.md (living: phases, ledger, next action) — for a new session to pursue autonomously across compactions; activatable directly by Codex goal mode; never pursues the goal itself
doc-to-html a standalone dark-themed HTML page rendered from a markdown report / audit / findings doc (TOC, keyboard nav, evidence appendix, print stylesheet), with a rigid editing discipline for later revisions
claim-check an unbiased, evidence-grounded investigation of a premise (ticket / hunch / question) against the current repo — a validity verdict with evidence plus a readiness dossier (or exactly what's missing); never implements the work
qa-sweep a team-scale QA pass over a decomposable surface — fans one agent per slice against the real running artifact, reproduces every verdict-moving finding firsthand before it counts, separates regressions from pre-existing bugs, and returns a verdict-first, confidence-tagged report; never fixes what it finds
empirical-proof a runtime proof of a just-finished change — gates on the app genuinely running (health-checked, right build), fans probe scenarios out to subagents under a raw-evidence contract (real MCP client connections and real HTTP only, never mocks or in-process harnesses), corroborates firsthand, and returns a verdict-first verified / broken / blocked report; never fixes local setup or the bugs it finds
code-quality-review an unusually strict, structure-first maintainability review over a branch's diff — hunts for "code judo" reframes that delete complexity, treats file-size explosions, spaghetti-branch growth, boundary leaks, and unearned abstractions as presumptive blockers, and prefers a few high-conviction structural findings over cosmetic nits
get-pr-comments a severity-grouped, prioritized action list from the active PR's conversation / review / inline comments, plus the open questions — read-only; never replies to, resolves, or reacts to a comment unless you explicitly ask
ui-demo-video a Playwright-driven video walkthrough of the running app after UI work — a PNG frame per scene the model reads to verify the UI itself (fix and re-record until the frames show the expected result), plus a webm/mp4 for the PR description; bundles the recording harness (scripts/harness.mjs, copied into the project)
route-work a three-line routing recommendation (route: / why: / grades:) for a task about to be dispatched — grades it on repo precedent, ambiguity, failure visibility, taste surface, and blast radius against the canonical model × effort table (kept in the skill as its single source of truth), and names the process pattern (direct dispatch / plan-review checkpoint / judge loop / taste pass); recommends only, never dispatches

Not included

No MCP servers or hooks. The agent-workshop-onboard skill and the profile-dependent or edit-capable agents (doc-indexer, wiki-maintainer, visual-implementer, research) live in the separate agent-workshop plugin, which adopts the full scaffold into a project.