Fifty-five portable workflows for Claude Code, Codex, OpenCode 1 and 2, and Pi.
Code review, security audit, feature development, frontend design, project memory, and authoring tools.
Install · Documentation · OpenCode 2 · Select Skills · Claude Code · Codex Quick Install · Pi Quick Install · Skills · Sandbox Contract · Invocation · Architecture · Credits
## OpenCode 2 and humanizer The current `main` branch supports OpenCode 2 and keeps OpenCode 1.18.7 support. Both hosts load all 55 skills and the three feature-development specialists. Smoke tests cover OpenCode 1.18.7, 2.0.14, and 2.0.32, including effective read-only permissions on V2. Install the current plugin in OpenCode 2: ```bash opencode plugin add "opencode-power-pack@git+https://github.com/waybarrios/opencode-power-pack.git" ``` The new `humanizer` skill rewrites English and Spanish copy in a direct, natural voice. It keeps facts, names, numbers, and code intact. The `authoring` profile now includes `humanizer`, `mcp-builder`, `paper-summarizer`, and `skill-creator`. Invoke it with `$humanizer` in Codex, `/humanizer` in OpenCode, or `/opencode-power-pack:humanizer` in Claude Code. **Use the Git installation for these additions.** The published npm package and tag `v0.5.0` still contain 54 skills and the V1-only plugin. They do not include `humanizer` or the OpenCode 2 plugin port. The npm selective installer can still copy its released skills into `.agents/skills` for OpenCode 2. For host commands, update steps, and the current catalog, see the [documentation](https://skills.waybarrios.com/docs/) and [skill reference](https://skills.waybarrios.com/docs/skills/). The [OpenCode compatibility guide](docs/opencode-compatibility.md) covers source selection and V2 permissions. ## Native sandboxes in v0.5.0 OpenCode Power Pack now includes an opt-in, fail-closed sandbox runner for safer command execution from **Codex, OpenCode, Claude Code, and Pi**. It gives all four agents one portable capability contract with four least-privilege profiles, while using native Seatbelt isolation on macOS and Bubblewrap on Linux. - **Lightweight:** one contained command tree, with no container image, daemon, or persistent sandbox to manage. - **Portable:** the same CLI and skill policy work from all four supported coding agents. - **Least privilege:** skills default to `observe`, `develop`, `network-read`, or `publish`, with explicit escalation checks. - **Fail closed:** unavailable isolation, undeclared escalation, missing confirmation, or backend failure stops execution instead of retrying outside the sandbox. Install the CLI, verify the boundary from the active agent session, and run a command with the skill's trusted profile: ```bash npm install --global @waybarrios/opencode-power-pack@0.5.0 opencode-power-pack sandbox doctor --json opencode-power-pack sandbox exec --skill code-review -- git status --short ``` The current guarantee is intentionally precise: the command and its descendants are `shell-contained`. Automatic host routing and whole-agent isolation are not claimed yet. See [run the sandbox from each agent](#run-the-sandbox-from-each-agent) and the [complete compatibility specification](docs/sandbox-compatibility.md). ## Claude Code Quick Install Run these commands inside Claude Code: ```text /plugin marketplace add waybarrios/opencode-power-pack /plugin install opencode-power-pack@opencode-power-pack ``` Run `/reload-plugins` if the installation summary requests it. Skills are namespaced to avoid collisions, so explicit invocations use names such as `/opencode-power-pack:code-review` and `/opencode-power-pack:feature-dev`. ## Codex Quick Install ```bash codex plugin marketplace add waybarrios/opencode-power-pack --ref main codex plugin add opencode-power-pack@opencode-power-pack ``` Start a new Codex session, open `/plugins` to confirm the installation, or invoke a workflow explicitly with `$code-review`, `$feature-dev`, `$security-review`, and the other bundled skills. ## Pi Quick Install ```bash pi install git:github.com/waybarrios/opencode-power-pack ``` Pi discovers the fifty-five skills declared by the package. Use `pi list` to verify the installation. For a project-local installation recorded in `.pi/settings.json`, add `-l`: ```bash pi install git:github.com/waybarrios/opencode-power-pack -l ``` ## Selective Install With npm Install only the workflows you want into the shared `.agents/skills` location recognized by Codex, OpenCode, and Pi: Use `npx` when you want to run the latest installer without keeping a global copy: ```bash # Inspect every profile and skill npx @waybarrios/opencode-power-pack list # Install the balanced recommended profile for your user npx @waybarrios/opencode-power-pack install --profile recommended # Install individual skills npx @waybarrios/opencode-power-pack install code-review security-review # Install a profile only for the current repository npx @waybarrios/opencode-power-pack install --profile review --project # Preview a selection without writing anything npx @waybarrios/opencode-power-pack install --profile security --dry-run # Update previously copied skills from the latest package npx @waybarrios/opencode-power-pack@latest install --profile review --force ``` Or install the package globally and use its shorter executable: ```bash npm install --global @waybarrios/opencode-power-pack opencode-power-pack list opencode-power-pack install --profile recommended opencode-power-pack install code-review security-review ``` To update or remove the global installation: ```bash npm update --global @waybarrios/opencode-power-pack npm uninstall --global @waybarrios/opencode-power-pack ``` The npm package is scoped as `@waybarrios/opencode-power-pack` to give it an unambiguous registry identity. The package still installs the shorter `opencode-power-pack` executable for direct CLI use. No global installation is required when using `npx`. ### GitHub Packages mirror Every GitHub release can also be published from its immutable tag to the repository's GitHub Packages page. npmjs remains the recommended public installation source because it works without GitHub authentication. To install the GitHub Packages mirror, first authenticate with a GitHub personal access token (classic) that has `read:packages`, then select the GitHub registry explicitly: ```bash npm login --scope=@waybarrios --auth-type=legacy --registry=https://npm.pkg.github.com npm install --global @waybarrios/opencode-power-pack --registry=https://npm.pkg.github.com opencode-power-pack list ``` Run the installer directly from the mirror after authentication: ```bash npx --registry=https://npm.pkg.github.com @waybarrios/opencode-power-pack list ``` Maintainers publish the mirror with the `Publish GitHub Package` workflow. Release events publish automatically, while a manual run requires an existing release tag such as `v0.5.0`. The workflow verifies that the tag matches `package.json`, uploads the artifact already verified by required release CI without rerunning lifecycle scripts, and authenticates with the repository-scoped `GITHUB_TOKEN`; no package token is stored in the repository. | Profile | Intended use | |---|---| | `recommended` | Balanced software development without the specialized security and ML catalogs | | `review` | Comprehensive and focused code review plus quality and differential review | | `feature-dev` | End-to-end feature workflow with its explorer, architect, and reviewer dependencies | | `frontend` | Frontend implementation and anti-generic design critique | | `security` | Security review, threat modeling, static analysis, validation, and reporting | | `huggingface` | Model discovery, local inference, training, evaluation, Spaces, and AWS deployment | | `authoring` | Skills, MCP servers, technical-paper summaries, and natural copy editing on `main` | | `project-memory` | Audit and maintain durable project guidance | Profiles automatically include required companion skills and can be combined with individual skill names. The default destination is `~/.agents/skills`; `--project` finds the current Git root and installs into its `.agents/skills` directory. Existing directories are skipped unless `--force` is provided, and replacements are staged with rollback plus per-skill locking. Every copied skill retains its provenance, third-party notices, and license texts. Use `--dry-run` to preview or `--all` to install the complete catalog. The commands above use the published npm catalog. To selectively install `humanizer` or the four-skill `authoring` profile from current source: ```bash npx --package=git+https://github.com/waybarrios/opencode-power-pack.git opencode-power-pack install humanizer --dry-run npx --package=git+https://github.com/waybarrios/opencode-power-pack.git opencode-power-pack install humanizer ``` Use `install --profile authoring` instead of `install humanizer` for all four authoring skills. These commands install into `~/.agents/skills`. For a project-local install, run the same command from the target repository with `--project`. The npm installer and the full Claude Code/Codex/OpenCode plugins are alternative activation paths. Claude Code users should prefer the namespaced plugin because the selective installer currently targets the portable `.agents/skills` location used by Codex, OpenCode, and Pi. If a full plugin is already active, selectively copying the same skills does not reduce that plugin's loaded catalog. ## Sandbox Capability Contract The package assigns every bundled skill a default least-capability profile and an explicit set of allowed escalations. Native plugin and package installations expose the skills but do not install the sandbox executable, so install the npm CLI in the same operating-system environment as the agent: ```bash npm install --global @waybarrios/opencode-power-pack@0.5.0 opencode-power-pack sandbox doctor opencode-power-pack sandbox resolve --skill code-review opencode-power-pack sandbox resolve --sandbox-profile publish --json opencode-power-pack sandbox exec --skill code-review -- rg TODO . ``` | Sandbox profile | Workspace | Network and credentials | External side effects | |---|---|---|---| | `observe` | Read | Denied | Denied | | `develop` | Write | Denied | Denied | | `network-read` | Write | Explicitly approved | Denied | | `publish` | Write | Explicitly approved | Explicit confirmation required | Run an allowed network escalation by naming the destination explicitly: ```bash opencode-power-pack sandbox exec \ --skill code-review \ --sandbox-profile network-read \ --allow-domain api.github.com \ -- curl https://api.github.com/repos/owner/repository ``` The literal `--` is required. Everything after it is the contained child command and its arguments. The runner preserves the child exit code and never falls back to an unsandboxed execution. ### Run the sandbox from each agent Run `sandbox doctor` inside the same agent session that will execute the command. Success in a separate terminal does not prove the active agent can start the native backend. | Host | How to invoke the runner | |---|---| | Codex | Ask Codex to use its shell tool for `opencode-power-pack sandbox doctor`, then for the complete `sandbox exec` command. | | OpenCode | Ask OpenCode to use its shell tool for `opencode-power-pack sandbox doctor`, then for the complete `sandbox exec` command. | | Claude Code | Ask Claude Code to use Bash for `opencode-power-pack sandbox doctor`, then for the complete `sandbox exec` command. | | Pi | Ask Pi to use its shell tool for `opencode-power-pack sandbox doctor`, then for the complete `sandbox exec` command. | Use this prompt in Codex, OpenCode, or Pi: ```text Use the shell tool to run `opencode-power-pack sandbox doctor`. If it succeeds, run `opencode-power-pack sandbox exec --skill code-review -- git status --short`. Do not run the child command separately and do not retry without the sandbox. ``` For Claude Code, use the same prompt with `Use Bash` in the first sentence. Keep the entire `sandbox exec` invocation in one host shell call. ### Platform requirements The runner requires Node.js 20.11.0 or newer. | Platform | Status | Additional requirements | |---|---|---| | macOS | Supported | Native Seatbelt; `sandbox doctor` performs an operational preflight. | | Linux | Supported | Trusted system installations of `bubblewrap`, `socat`, and `rg`, plus working unprivileged user and network namespaces. | | WSL2 | Expected, dedicated CI pending | The same Linux dependencies and namespace support. | | WSL1 | Unsupported | Required namespace primitives are unavailable. | | Windows native | Fails closed | Platform-specific provisioning and boundary tests are not complete. | The contract continues to report `advisory` because skill metadata does not enforce itself. The npm package also provides an opt-in `sandbox exec` runner backed by a pinned native runtime. When its operational preflight succeeds, `sandbox doctor` reports `shell-contained`; `Strict ready` remains `no` until a host adapter blocks ordinary tool bypasses. For `network-read`, the runner enforces conventional HTTP read methods (GET, HEAD, and OPTIONS). This is a best-effort side-effect boundary because an approved endpoint can still implement unsafe GET behavior. Use narrow destination grants and expose only credentials required for the retrieval. The runner is lightweight and on demand: it contains one command tree with native OS primitives, starts no persistent container or daemon, denies unrelated host reads, and deletes its private temporary directory after execution. `sandbox doctor` exits nonzero when the native boundary is unavailable. See [Sandbox Compatibility Across Coding Agents](docs/sandbox-compatibility.md) for the current Codex, Claude Code, OpenCode, and Pi matrix, platform requirements, activation design, fail-closed behavior, and adapter roadmap. Selective installations include a generated `SANDBOX_POLICY.json` beside each copied `SKILL.md`. The file is self-describing and preserves the skill's default profile even when the central package is not present. It remains advisory until an enforcement backend and host adapter are active. ## Why This Exists Claude Code, Codex, OpenCode, and Pi read `SKILL.md` workflows, but many valuable workflows originated as Claude-specific commands and agents. Copying those artifacts directly does not preserve their orchestration, permissions, or subagent behavior. This package exposes the shared skills as a namespaced Claude Code plugin, registers feature-development specialist roles as read-only OpenCode subagents, lets Codex execute the same phase assignments with its native subagent workflow, and exposes all fifty-five skills as a Pi package. It ships immutable provenance for every upstream work. It complements [obra/superpowers](https://github.com/obra/superpowers), which provides process skills such as brainstorming, TDD, debugging, and plan execution. ## What's Inside | Category | Skill | Purpose | |---|---|---| | Review | `code-review` | Multi-agent PR review with confidence filtering and reproduction scenarios | | Review | `security-review` | Security-focused review with category coverage and exploit-path validation | | Review | `supply-chain-risk-auditor` | Flag dependencies at heightened risk of exploitation or takeover | | Review | `sharp-edges` | Identify error-prone APIs and footgun configurations before they cause security mistakes | | Review | `insecure-defaults` | Detect fail-open insecure defaults (hardcoded secrets, weak auth, permissive config) | | Review | `fp-check` | Systematically verify a suspected security bug to a TRUE/FALSE POSITIVE verdict | | Review | `vuln-report` | Draft a single-vulnerability disclosure report in GitHub advisory style | | Review | `agentic-actions-auditor` | Audit GitHub Actions workflows for prompt-injection risk in AI agent integrations | | Review | `security-threat-model` | Model threats using repository evidence, trust boundaries, assets, and abuse paths | | Review | `differential-review` | Security-focused differential review of a diff/PR with blast-radius analysis | | Review | `variant-analysis` | Find similar vulnerabilities/bugs across a codebase from one initial pattern | | Review | `sarif-parsing` | Parse, filter, dedupe, and convert SARIF output from CodeQL/Semgrep/other scanners | | Review | `semgrep` | Run a Semgrep static analysis scan with approval-gated ruleset selection | | Review | `semgrep-rule-creator` | Write custom Semgrep rules for a specific vulnerability/bug pattern | | Review | `semgrep-rule-variant-creator` | Port an existing Semgrep rule to additional target languages | | Review | `codeql` | Run a CodeQL scan using interprocedural data flow and taint tracking | | Feature development | `feature-dev` | Seven-phase workflow from discovery through implementation and review | | Feature development | `code-explorer` | Trace a feature across entry points, layers, state, and dependencies | | Feature development | `code-architect` | Produce a file-level architecture and implementation blueprint | | Feature development | `code-reviewer` | Adversarial review of a focused local change set | | Design | `frontend-design` | Create distinctive interfaces with an accessibility and craft rubric | | Design | `ai-slop` | Rubric to judge whether a design is generic "AI slop" or genuinely product-fit | | Code quality | `code-quality` | Linting, complexity, and review checklists across Rust, TypeScript, Python, and shell | | Code quality | `design-patterns` | Pattern trade-offs (when to use, when to skip) for Rust, TS/React, and Django/Python | | Research | `paper-summarizer` | Extract actionable findings and a claim-evidence map from academic/technical papers | | Authoring | `humanizer` | Rewrite AI-generated copy in ES/EN into direct, natural prose without inventing facts | | Authoring | `mcp-builder` | Design and build MCP servers in TypeScript or Python | | Authoring | `skill-creator` | Create, test, and improve reusable `SKILL.md` workflows | | Hugging Face | `hf-cli` | Use the `hf` CLI for auth, cache, repos, jobs, papers, and Spaces | | Hugging Face | `hf-mem` | Estimate memory needed to load Safetensors/GGUF weights for inference | | Hugging Face | `huggingface-best` | Find and compare the best model for a task via official benchmark leaderboards | | Hugging Face | `huggingface-datasets` | Dataset Viewer API workflows for metadata, pagination, search, and filters | | Hugging Face | `huggingface-papers` | Look up and analyze Hugging Face / arXiv paper pages via the papers API | | Hugging Face | `huggingface-paper-publisher` | Publish and manage research papers on the Hub, including research-article templates | | Hugging Face | `huggingface-tool-builder` | Build reusable scripts/tools around the Hugging Face API | | Hugging Face | `huggingface-local-models` | Select and run GGUF models locally with llama.cpp (CPU/Metal/CUDA/ROCm) | | Hugging Face training | `huggingface-llm-trainer` | Fine-tune LLMs (TRL: SFT/DPO/GRPO, or Unsloth) on Hugging Face Jobs | | Hugging Face training | `huggingface-vision-trainer` | Fine-tune object detection, image classification, and SAM/SAM2 segmentation models | | Hugging Face training | `train-sentence-transformers` | Train bi-encoder/cross-encoder/sparse embedding models | | Hugging Face training | `trl-training` | Train with the TRL CLI using SFT, DPO, GRPO, KTO, RLOO, and reward modeling | | Hugging Face training | `huggingface-community-evals` | Run local evaluations with inspect-ai / lighteval | | Hugging Face training | `huggingface-trackio` | Track and visualize ML training experiments with Trackio | | Hugging Face Spaces | `huggingface-spaces` | Build, deploy, and debug Gradio/Docker/static Spaces | | Hugging Face Spaces | `huggingface-gradio` | Build Gradio web UIs and demos in Python | | Hugging Face Spaces | `huggingface-lora-space-builder` | Build and publish a Gradio Space demo for a user-provided LoRA | | Hugging Face Spaces | `huggingface-zerogpu` | Write/debug ZeroGPU-constrained Gradio Space code | | Hugging Face cloud | `hf-cloud-aws-context-discovery` | Discover the user's local AWS context before any AWS work | | Hugging Face cloud | `hf-cloud-python-env-setup` | Set up an isolated Python env for SageMaker/AWS work | | Hugging Face cloud | `hf-cloud-sagemaker-deployment-planner` | Entry point for planning a SageMaker model deployment | | Hugging Face cloud | `hf-cloud-sagemaker-iam-preflight` | Ensure a usable SageMaker execution role exists before deploying | | Hugging Face cloud | `hf-cloud-serving-image-selection` | Pick the right SageMaker serving container image | | Hugging Face cloud | `hf-cloud-sagemaker-production-defaults` | Create a SageMaker endpoint with autoscaling, alarms, and tagging by default | | JavaScript | `transformers-js` | Run Hugging Face models directly in JS/TS (browser or Node/Bun/Deno) | | Project memory | `agents-md-improver` | Audit project rules and propose targeted improvements | | Project memory | `agents-md-revise` | Capture durable session learnings in project rules | `code-explorer`, `code-architect`, and `code-reviewer` are standalone skills and specialist roles used by `feature-dev`. OpenCode registers named least-privilege agents; Codex can carry out the same assignments with native subagents and can use matching custom agents when the user configures them. ### Which Code Review Skill? | Situation | Use | Why | |---|---|---| | PR, branch, commit range, or complete pending diff | `code-review` (Comprehensive Code Review) | Freezes the scope, runs multiple review lenses, cross-checks candidates, and can post validated findings to GitHub when requested | | One file, function, or small local change | `code-reviewer` (Focused Code Review) | Uses a lighter two-pass review with full-file and caller context | | `feature-dev` quality phase | `code-reviewer` (Focused Code Review) | Acts as the focused specialist role dispatched by the feature workflow | They are complementary rather than duplicate workflows. Codex shows distinct user-facing titles, Claude Code namespaces both under `opencode-power-pack`, and the stable skill IDs continue to work across every host. Start with `code-review` for merge decisions and `code-reviewer` for focused iteration. ## Installation ### Prerequisites - Git - Node.js 20.11.0 or newer, including npm/npx, for [selective installation](#selective-install-with-npm) and the sandbox runner - One supported host: - A current Claude Code installation with plugin support: