Build fast without breaking things: a new product on a deadline, a React front end, a release that has to hold, a launch, or a fleet of backend services moved to .NET 10. The kit stops the agent before the costly mistakes, and gives it the playbooks and the specialists for everything else.

$ npx @rxova/agent-kit install --profile hackathon
Installed rx-ai (hackathon: 15 agents, 24 skills, 10 hooks) for Claude Code into ~/.claude.

Then, in any session, when the agent reaches for a shortcut:

● Bash(git commit --no-verify -m "wip")
  ⎿  rx-ai no-bypass: git commit is set to skip its hooks. Run it without the bypass;
     if a hook fails, fix what it reports.

● Bash(pnpm dev)
  ⎿  rx-ai dev-server: This starts a process that never exits. Run it with
     run_in_background: true and read its output from there.

What you get

| | | | ---------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | 🛡️ 6 guards | Block the mistakes that cost an afternoon: skipping git hooks, AI attribution in commits, destroying data or work, leaking keys, hanging on a dev server, loosening lint rules. | | 🔧 4 helpers | Format and lint every edit, save a snapshot before compaction, point a new session at the last handoff, and warn before the context fills up. | | 📚 51 skills | Playbooks Claude loads when a request matches: workflow from kickoff to demo, stacks, platforms, React, QA, marketing, the .NET 10 upgrade, Datadog. | | 🤖 25 agents | Specialists Claude hands work to, each with the fewest tools its job needs and a model sized to it. | | 📦 Profiles | Install a set, not everything: every installed description is loaded into every session. | | ↩️ Clean uninstall | A manifest records every file and hook the install wrote. Uninstall removes exactly those; nothing you wrote is touched. |

Install

npx @rxova/agent-kit install                                 # core: the guards and the everyday set
npx @rxova/agent-kit install --profile hackathon             # everything for building a product fast
npx @rxova/agent-kit install --profile react                 # React: Redux Toolkit, hooks, code splitting, performance
npx @rxova/agent-kit install --profile qa                    # test plans, exploratory runs, bug reports, flaky tests
npx @rxova/agent-kit install --profile marketing             # positioning, landing copy, launch, CEO view, in-app voice
npx @rxova/agent-kit install --profile dotnet --target both  # .NET and Datadog, for Claude Code and OpenCode
npx @rxova/agent-kit install --profile full                  # all of it
npx @rxova/agent-kit list                                    # every item, and which profiles include it

Restart Claude Code afterwards so it loads the new agents, skills and hooks. Run install again to update or widen: with no --profile it keeps the last selection, so --add rx-kickoff alone adds one item.

Profiles

With no --profile, you get core: the guards, and the agents and skills used on every change. The headline numbers above are full.

| Profile | Agents | Skills | Hooks | For | | --------------- | :----: | :----: | :---: | --------------------------------------------------------------------------------------- | | core | 6 | 5 | 6 | Every repository. The guards, plan → build → review → ship. | | hackathon | 15 | 24 | 10 | A product on a deadline. Everything but rx-tdd, rx-theme-audit and the sets below. | | react | 10 | 15 | 10 | React front ends. core, plus every rx-fe-* item and the UI and test helpers. | | qa | 10 | 15 | 10 | Quality assurance. core, plus every rx-qa-* item and the test and security helpers. | | marketing | 12 | 12 | 10 | Running the product. core, plus every rx-mkt-* item, the pitch, demo and docs. | | dotnet | 13 | 18 | 10 | Service migrations. core, plus .NET, EF Core, Datadog and helpers. | | full | 25 | 51 | 10 | Everything. |

• core: the 6 guards; rx-planner, rx-reviewer, rx-scout, rx-builder, rx-debugger, rx-build-fixer; rx-verify, rx-handoff, rx-slice, rx-debug, rx-ship.
• dotnet: core, plus every .NET, EF Core and Datadog skill and agent, plus rx-architect, rx-test-writer, rx-security, rx-researcher, rx-parallel, rx-tdd, rx-security-sweep, rx-postgres, rx-deploy-container and the 4 helper hooks.
• react: core, plus every rx-fe-* skill and agent, plus rx-ui, rx-test-writer, rx-react-web, rx-ui-kit, rx-e2e, rx-tdd and the 4 helper hooks.
• qa: core, plus every rx-qa-* skill and agent, plus rx-test-writer, rx-security, rx-e2e, rx-tdd, rx-theme-audit, rx-security-sweep and the 4 helper hooks.
• marketing: core, plus every rx-mkt-* skill and agent, plus rx-pitch, rx-doc-writer, rx-researcher, rx-demo, rx-kickoff and the 4 helper hooks.

rxova-agent-kit list --profile <name> prints any profile item by item.

How it works

You install once, then talk to Claude as usual. The kit works at three levels:

| Part | Who starts it | How | | ------------- | ------------------------------------- | ----------------------------------------------------------------------------------------------------- | | 🪝 Hooks | Claude Code, on its own | Fire on events: before a command or edit, after an edit, at session start and end, before compaction. | | 📚 Skills | Claude when a request matches, or you | Claude loads a skill whose description fits. Type /rx-kickoff, /rx-demo, … to load one yourself. | | 🤖 Agents | Claude when a task matches, or you | Claude delegates to an agent and gets a summary back. Say "have rx-reviewer check this" to pick one. |

A guard only watches Claude's own tool calls. When it blocks something you really want, run it yourself in a terminal, or switch that guard off for the session (see Settings).

1. "We're building a split-the-bill app for group trips, 24 hours. Let's kick off." Claude follows rx-kickoff: the demo moment, one golden path, a boring stack, a CLAUDE.md, a deploy in the first hour. rx-architect designs the data model; rx-slice cuts the work into slices and stops for your go-ahead.
2. "Go, slice 1." rx-builder builds it. After every edit quick-check lints the file. pnpm dev in the foreground is stopped by dev-server; a Supabase key in a source file by secret-guard; a --no-verify commit by no-bypass. rx-ship opens the pull request.
3. Hour 14. context-nudge says the context is 60% full; Claude writes a handoff note (rx-handoff). memory-snapshot saves state before the compact, and next session handoff-reminder points Claude at it.
4. "Login works locally but not on Vercel." rx-debug and rx-deploy-vercel. A panicked git reset --hard on uncommitted work is stopped by danger-zone.
5. "Get us demo-ready." rx-security-sweep and rx-security check for leaks and missing row-level security, rx-demo adds seed data and a click-by-click script, rx-pitch writes the pitch.

1. "Upgrade the orders service from .NET 6 to .NET 10." rx-dotnet-upgrade: an inventory of projects, packages, images and CI; a green baseline; then the SDK, target framework and packages. rx-dotnet-migrator works through breaking changes one area at a time, with a log.
2. Tests. rx-dotnet-testing detects the framework and mocking library, converts FluentAssertions to Shouldly, and adds WebApplicationFactory and Testcontainers tests where behaviour changed. dotnet run in the foreground is stopped by dev-server, dotnet ef database drop by danger-zone.
3. Observability. rx-datadog-dotnet checks the tracer supports .NET 10 and the service tags still line up; after the deploy, rx-observability compares error rate and latency before and after through the Datadog MCP server, and rx-datadog-monitors fixes monitors on renamed resources.
4. Review and ship. rx-dotnet-reviewer checks async and DI mistakes, EF Core query and migration safety, and behaviour the framework change altered; rx-ship opens the pull request.

Hooks

Every hook runs through one bundled file, rx-ai/hooks.js, that imports only Node: no npx per call, about 80 ms each. A guard that blocks exits 2 and tells the agent why and what to do instead. On anything unexpected a hook lets the call through: it never breaks a session because of its own failure.

🛡️ Guards — in every profile

| Hook | Blocks | | ---------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | no-bypass | Git calls that skip the repository's hooks: --no-verify, commit -n, HUSKY=0, changing core.hooksPath. | | no-attribution | Commit messages and gh pr bodies that credit an AI assistant: Co-Authored-By: Claude, session trailers, "Generated with Claude", 🤖. Message and body files included. | | danger-zone | rm -r outside the project, force pushes to main, reset --hard on a dirty tree, SQL DROP/TRUNCATE, database resets, terraform destroy and other teardown. | | dev-server | Dev servers and watchers in the foreground (pnpm dev, vite, uvicorn, dotnet watch, docker compose up, …); says to use run_in_background. | | config-lock | Edits to an existing lint, format, type, commit or coverage config. Creating one is fine; Directory.Build.props and Directory.Packages.props stay editable. | | secret-guard | API keys, tokens and private keys written into source files (AWS, Anthropic, OpenAI, Stripe, GitHub, Slack, Google, Datadog, Azure, Supabase). .env files pass. |

🔧 Helpers — in hackathon, dotnet and full

| Hook | When | Does | | ------------------ | ------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------ | | quick-check | after Edit/Write | Formats and lints the edited file with the project's own tools (Biome, Prettier + ESLint, Ruff, dotnet format) and reports back. | | memory-snapshot | before compaction, session end | Saves branch, uncommitted files, recent commits and the last requests to .claude/handoff/auto/, git-ignored, keeping the last ten. | | handoff-reminder | session start | Points the agent at the newest handoff note (a week at most) or snapshot (two days). | | context-nudge | after any tool | At 60% and 80% of the context window, tells the agent to reach a break, write a handoff and suggest /compact. |

Settings

| Variable | Example | Effect | | ---------------------- | ----------------------- | ------------------------------------------------------------------------- | | RX_AI_OFF | RX_AI_OFF=config-lock | Switches hooks off, e.g. for one session: RX_AI_OFF=config-lock claude. | | RX_AI_CONTEXT_WINDOW | 1000000 | The window size context-nudge measures against (200,000 by default). |

Status line

--statusline also sets Claude Code's status line to the kit's script, so a new machine gets the same bar from the same install:

npx @rxova/agent-kit install --statusline      # add it (kept on later installs)
npx @rxova/agent-kit install --no-statusline   # take it out again

Two lines: the model, a context bar, the project (and worktree) and git branch with staged, modified, untracked and ahead/behind counts; then session cost and time, lines changed, 5-hour and weekly plan usage, and the output style and Claude Code version. It needs bash, jq and git, and a Powerline or Nerd Font for the separators. A status line someone else set is left alone unless you add --force; uninstall removes only the kit's.

Skills

Each skill is a folder with a SKILL.md: when to use it, the steps, and an example.

| Group | Skills | | ---------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | 🧭 Workflow | rx-kickoff · rx-timebox · rx-slice · rx-parallel · rx-tdd · rx-debug · rx-verify · rx-ship · rx-handoff · rx-e2e · rx-security-sweep · rx-demo · rx-theme-audit | | 🧱 Stacks | rx-react-web · rx-ui-kit · rx-node-api · rx-python-api · rx-expo · rx-claude-api · rx-auth | | ☁️ Platforms | rx-supabase · rx-postgres · rx-deploy-vercel · rx-deploy-cloudflare · rx-deploy-container · rx-aws | | ⚛️ React | rx-fe-redux-toolkit · rx-fe-code-splitting · rx-fe-hooks · rx-fe-performance · rx-fe-state · rx-fe-testing | | 🧪 QA | rx-qa-test-plan · rx-qa-exploratory · rx-qa-bug-report · rx-qa-regression · rx-qa-flaky · rx-qa-a11y | | 📣 Marketing | rx-mkt-positioning · rx-mkt-landing · rx-mkt-launch · rx-mkt-ceo · rx-mkt-voice | | 🟣 .NET | rx-dotnet-upgrade · rx-dotnet-api · rx-efcore · rx-dotnet-workers · rx-dotnet-testing | | 🐶 Datadog | rx-datadog-dotnet · rx-datadog-investigate · rx-datadog-monitors |

| Skill | Use it to | | ------------------------ | ------------------------------------------------------------------------------------------------------ | | rx-kickoff | Turn an idea into a scoped plan, a repo with agent instructions, and a live deploy in hour one. | | rx-timebox | Run the build against the clock: checkpoints, cut lists, a feature freeze, the last two hours. | | rx-slice | Break a feature into thin end-to-end slices and ship them one at a time. | | rx-parallel | Run several sessions and agents at once with worktrees, without collisions. | | rx-tdd | Test first where it pays, and skip it where it does not. | | rx-debug | Reproduce, isolate and fix a bug at its root, with a regression test. | | rx-verify | Run the repository's own gate before calling work done. | | rx-ship | Get a change merged the repository's way: branch, commits, checks, pull request. | | rx-handoff | Write a note a fresh session can resume from, or resume from one. | | rx-e2e | Put a Playwright smoke suite on the demo path. | | rx-security-sweep | Do a 30-minute security pass before the demo. | | rx-demo | Make the demo impossible to fail: seed data, a script, fallbacks, a backup video. | | rx-theme-audit | Measure a site's dark and light themes in a browser and trace each problem to its source. | | rx-react-web | Build with Next.js or Vite: rendering, data fetching, forms, env vars. | | rx-ui-kit | Get a good-looking, accessible UI fast with Tailwind and shadcn/ui, dark mode included. | | rx-node-api | Build a typed API with Hono (or Fastify/Express): validation, errors, auth, CORS, tests. | | rx-python-api | Build an API with FastAPI and uv: models, dependencies, async database, tests, Docker. | | rx-expo | Ship a React Native app with Expo: routing, devices, env vars, auth, EAS. | | rx-claude-api | Add AI features with the Claude API: streaming, tools, structured output, caching, cost caps. | | rx-auth | Pick and wire authentication fast, with a seeded demo account. | | rx-supabase | Use Supabase: local dev, migrations, row-level security, storage, edge functions, types. | | rx-postgres | Use Postgres with Drizzle or Prisma: schema, migrations, pooling, indexes, seeds. | | rx-deploy-vercel | Deploy to Vercel or Netlify: previews, env vars, monorepos, limits, rollback. | | rx-deploy-cloudflare | Deploy Workers and Pages with wrangler: bindings, secrets, D1, R2, KV. | | rx-deploy-container | Deploy containers to Fly.io, Railway or anywhere Docker runs. | | rx-aws | Take the fast paths on AWS, with a budget alarm first and a teardown list last. | | rx-fe-redux-toolkit | Use Redux Toolkit 2: store, slices, typed hooks, RTK Query tags and optimistic updates, migration. | | rx-fe-code-splitting | Split by route and feature, preload on intent, recover from stale chunks, hold a bundle budget. | | rx-fe-hooks | Write custom hooks, drop effects you don't need, get deps and cleanup right, use React 19 hooks. | | rx-fe-performance | Profile first, cut re-renders, lean on the React Compiler, virtualise lists, fix LCP, INP and CLS. | | rx-fe-state | Decide where state lives: local, URL, server cache, store, context or form. Derive, don't duplicate. | | rx-fe-testing | Test components with Testing Library role queries, user-event and MSW; split component from e2e. | | rx-qa-test-plan | Plan a feature or release by risk: ranked risks, a test matrix, automated vs manual, exit criteria. | | rx-qa-exploratory | Run time-boxed exploratory sessions with charters, heuristics and tours; turn findings into tests. | | rx-qa-bug-report | Write bugs anyone can reproduce: minimal steps, evidence, severity vs priority, then a failing test. | | rx-qa-regression | Scope smoke and regression runs from the diff and the risks, with a release checklist and sign-off. | | rx-qa-flaky | Reproduce a flaky test with repeats and shuffles, fix the root cause, quarantine with an owner. | | rx-qa-a11y | Test accessibility: axe, keyboard, screen reader, 400% reflow, forms, mapped to WCAG 2.2 AA. | | rx-mkt-positioning | Pin down who it's for, the alternative, what's unique, the category and the one-liner. | | rx-mkt-landing | Write landing page and store listing copy, meta and OG text, in the site's own file, and measure it. | | rx-mkt-launch | Plan a launch: channels and their rules, assets, a day-of timeline, funnel events, a retro. | | rx-mkt-ceo | Think like the CEO: north-star and input metrics, weekly priorities, decision memos, pricing, updates. | | rx-mkt-voice | Set the app's voice and tone, fix microcopy and errors, keep a glossary and i18n-safe strings. | | rx-dotnet-upgrade | Move a service from .NET 6 or 8 to .NET 10: inventory, baseline, packages, breaking changes, CI. | | rx-dotnet-api | Build and upgrade ASP.NET Core APIs: minimal APIs or controllers, validation, ProblemDetails, OpenAPI. | | rx-efcore | Upgrade and use EF Core safely: breaking changes, migrations and production scripts, performance. | | rx-dotnet-workers | Run hosted services and messaging reliably: shutdown, scopes, retries, idempotency, outbox. | | rx-dotnet-testing | Test with Shouldly: detect the framework, convert FluentAssertions, WebApplicationFactory. | | rx-datadog-dotnet | Instrument a .NET service: tracer, service tags, log correlation, metrics, through an upgrade. | | rx-datadog-investigate | Go from an alert to the failing span and its logs, before and after a deploy. | | rx-datadog-monitors | Manage monitors, SLOs and dashboards as code, and keep them valid after renames. |

Agents

Each agent gets the fewest tools its job needs and a model sized to it. Read-only agents cannot edit, so a reviewer never "fixes" what it was asked to review.

| Agent | Model | Edits | Use it to | | -------------------- | :----: | :---: | ----------------------------------------------------------------------- | | rx-planner | fable | — | Turn a request into a sliced plan grounded in the code. | | rx-architect | fable | — | Design the system for a time-boxed build: parts, data, boundaries. | | rx-security | fable | — | Find what would embarrass the demo: secrets, auth gaps, RLS, injection. | | rx-reviewer | opus | — | Review a diff and report only defects it can back. | | rx-researcher | opus | — | Answer "how do I do X with Y" from official docs, with sources. | | rx-dotnet-reviewer | opus | — | Review a .NET diff: async, DI lifetimes, EF Core, upgrade behaviour. | | rx-observability | opus | — | Investigate production in Datadog and report with evidence. | | rx-builder | opus | ✓ | Build one slice end to end, with tests. | | rx-debugger | opus | ✓ | Find a bug's root cause and fix it with a regression test. | | rx-test-writer | opus | ✓ | Add tests that check behaviour. | | rx-build-fixer | opus | ✓ | Turn a red build green with the smallest honest diff. | | rx-ui | opus | ✓ | Build and polish UI: responsive, accessible, every state. | | rx-db | opus | ✓ | Design schemas, migrations, policies and demo seed data. | | rx-deployer | opus | ✓ | Ship to the chosen platform and prove the live URL works. | | rx-pitch | opus | ✓ | Write the demo script and the pitch. | | rx-dotnet-migrator | opus | ✓ | Upgrade one .NET service to .NET 10, with a per-service log. | | rx-fe-builder | opus | ✓ | Build React features in the repo's own conventions, and verify them. | | rx-fe-reviewer | opus | — | Review React: hooks, effects, re-renders, state, a11y, bundle size. | | rx-qa-tester | opus | ✓ | Explore the running app against a charter; report bugs with evidence. | | rx-qa-flake-hunter | opus | ✓ | Fix a flaky test at its root and prove it with repeat runs. | | rx-mkt-advisor | fable | ✓ | A CEO's second opinion: test plans against the goal, cut, write memos. | | rx-mkt-marketer | opus | ✓ | Write positioning, landing and launch copy from what the code does. | | rx-mkt-copywriter | opus | ✓ | Rewrite in-app copy to the voice guide; strings only, never logic. | | rx-scout | sonnet | — | Find where things live without flooding the context. | | rx-doc-writer | sonnet | ✓ | Write a README a newcomer can follow in five minutes. |

rx-observability has no tool list on purpose, so it can reach the Datadog MCP server; it changes nothing unless asked.

Commands

| Command | What it does | | --------------------------- | ------------------------------------------------------------------------------------------------------ | | rxova-agent-kit list | Every agent, skill and hook, the profiles that include it, and what it is for. | | rxova-agent-kit install | Installs a profile, or updates an existing install. | | rxova-agent-kit uninstall | Removes the files each install wrote and its hook entries in settings.json. Nothing else is touched. | | rxova-agent-kit status | Each installed target's profile and version, and anything missing or edited. Exits 1 when out of step. |

| Flag | Commands | Effect | | --------------- | -------------------------------- | --------------------------------------------------------------------------------------- | | --profile <p> | install, list | core (default on a first install), hackathon, dotnet, full. | | --target <t> | install, uninstall, status | claude (default on a first install), opencode or both. Later: the installed ones. | | --add a,b | install | Adds items on top of the profile. | | --skip c | install | Leaves items out of the profile. | | --statusline | install | Sets Claude Code's status line to the kit's script; --no-statusline takes it out. | | --project | install, uninstall, status | Works on ./.claude or ./.opencode instead of the user folder. | | --dry-run | install, uninstall | Prints the plan and writes nothing. | | --force | install | Overwrites files (and a status line) the kit did not write. |

| Target | User install | With --project | | ---------- | ----------------------------------------------------- | ---------------- | | claude | ~/.claude | ./.claude | | opencode | ~/.config/opencode (or $XDG_CONFIG_HOME/opencode) | ./.opencode |

Each target records what the install wrote in rx-ai/manifest.json. The install refuses to overwrite a file it did not write, and reads and checks settings.json before writing anything.

Mods

Three Claude Code mods live in mods/. They are Claude Code plugins, separate from the install bin and from the npm package: install each one from the rxova/shared marketplace, at the prompt of a terminal session.

/plugin install rx-attribution --marketplace rxova/shared
/plugin install rx-worktrees --marketplace rxova/shared
/plugin install rx-handoff-band --marketplace rxova/shared

Answer y to add the marketplace, then pick a scope.

| Mod | What it does | | ----------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | rx-attribution | Empties the commit trailer and the PR footer Claude Code asks the model to write. After gh pr create or gh pr edit, reads the PR body and removes any attribution line with gh pr edit. After git commit, flags an attribution trailer and, when authorEmail is set, a commit whose author or committer has another email. | | rx-worktrees | /worktrees opens a pane listing every worktree of the repository with its branch and pull request state. Pressing a row copies its path; Prune runs git worktree prune when a worktree's folder is gone. | | rx-handoff-band | Shows the newest .claude/handoff/yyyy-mm-dd*.md note above the prompt until the first prompt. Resume (h) fills the prompt to read it and continue. |

Options are set in /config: rx-attribution takes authorEmail (empty skips the author check) and fixPrBody (off only reports); rx-handoff-band takes maxDays (7).

Each mod is checked with claude plugin validate packages/agent-kit/mods/<mod> and claude plugin test packages/agent-kit/mods/<mod>.

OpenCode

--target opencode (or both) installs the same kit for OpenCode:

• Agents become OpenCode subagents (mode: subagent); the tools an agent was not given become a permission block. They run on the model of the agent that calls them.
• Skills use the same SKILL.md format. OpenCode also reads .claude/skills/, so with --target both they are written once.
• Hooks run through plugins/rx-kit.js, which calls the same runner: a guard fails the tool call with its message, quick-check appends its report to the edit's output, and handoff-reminder adds its pointer to the system prompt. apply_patch edits are checked file by file.

context-nudge and the session-end snapshot do not run there: OpenCode has no transcript file to measure and no session-end event. Restart OpenCode after installing.

Programmatic use

The package exports what the bin is built from: runHook, runHooks, hooks, catalog, profiles, selectItems, resolveTargets, planInstall, installCopies, hookGroups, withOwnHooks, withoutOwnHooks, opencodeAgent, opencodeHooks, opencodePlugin, and their types.

License

MIT