# opencode-power-pack: 54 Portable Skills for Codex, OpenCode, Claude Code and Pi > A pack of 54 workflow skills that installs into four different coding agents, plus an opt-in sandbox runner. Useful if you already use one of those agents and want repeatable review and security prompts without writing them yourself. **waybarrios/opencode-power-pack** — 54 rigorous skills for Codex, OpenCode, and Pi: code review, security audit, feature development, frontend design, MCP tools, Hugging Face ML/training, and more. - Repository: https://github.com/waybarrios/opencode-power-pack - Website: https://skills.waybarrios.com - Stars: 527 · Forks: 42 - Language: Python - License: MIT - Published: 2026-09-10 · Updated: 2026-09-10 · Language: en - Canonical page: https://hysenlabs.com/projects/waybarrios-opencode-power-pack ## The problem: the same review prompt, retyped in four agents Anyone who runs a coding agent on a real repository ends up writing the same instructions over and over. Review this diff for missing error handling. Check this dependency for known issues. Scaffold a feature with tests. Those prompts are personal, they drift between sessions, and they are not shared with the rest of the team. opencode-power-pack packages that work as files. The README describes fifty-four portable workflows covering code review, security audit, feature development, frontend design, project memory and authoring tools, and it targets four hosts at once: Claude Code, Codex, OpenCode and Pi. The intended audience is a developer who already has one of those agents installed and wants a fixed, inspectable set of procedures rather than an ad hoc prompt library. The portability claim is the interesting part. The same skill directory is consumed by four products with different plugin formats, which is why the repository ships separate .claude-plugin, .codex-plugin and .opencode directories at the top level alongside a pi key in package.json that points at ./skills. ## How the skills are packaged and discovered A skill here is not a service. It is a directory of instructions that the host agent reads and follows, and the repository is mostly that content: a skills directory, plus per-host manifests that tell each agent where to find it. The npm package declares its entry point as .opencode/plugins/opencode-power-pack.js and exposes a single executable, opencode-power-pack, mapped to bin/opencode-power-pack.mjs. The engines field sets the floor: Node >=20.11.0 and opencode >=1.18.7. That second constraint is unusual, because it means the npm package asserts a minimum version of a separate tool that is not its own dependency. If you are on an older OpenCode build, the package metadata says you are outside the supported range. Version 0.5.0 adds something structurally different from the skill files: an opt-in sandbox runner. According to the README, it uses native Seatbelt isolation on macOS and Bubblewrap on Linux, with no container image, daemon or persistent sandbox, and it exposes four least-privilege profiles named observe, develop, network-read and publish. The stated failure mode is fail-closed: unavailable isolation, an undeclared escalation, a missing confirmation or a backend failure stops execution rather than retrying outside the sandbox. The README is careful about the boundary, and this is worth repeating precisely: the current guarantee is that the command and its descendants are shell-contained, and automatic host routing and whole-agent isolation are explicitly not claimed. That is a narrower promise than the word sandbox usually implies, and the project says so itself. ## Installing it and running a first review There are four install paths, one per host, plus a selective npm installer. Start with the selective route if you want to see what you are getting before anything is written to disk. The list command prints every profile and skill, and the dry-run flag previews a selection without copying files. ```bash npx @waybarrios/opencode-power-pack list npx @waybarrios/opencode-power-pack install --profile security --dry-run ``` When the selection looks right, install the balanced profile for your user account, or name individual skills. The README shows code-review and security-review as the example pair. ```bash npx @waybarrios/opencode-power-pack install --profile recommended npx @waybarrios/opencode-power-pack install code-review security-review ``` Skills land in the shared .agents/skills location, which the README says is recognized by Codex, OpenCode and Pi. Adding --project scopes the install to the current repository instead of your user directory. To update previously copied skills from a newer package, the README gives --force with the @latest tag. ```bash npx @waybarrios/opencode-power-pack install --profile review --project npx @waybarrios/opencode-power-pack@latest install --profile review --force ``` Claude Code and Pi use their own mechanisms. Inside Claude Code the two commands are /plugin marketplace add waybarrios/opencode-power-pack and /plugin install opencode-power-pack@opencode-power-pack, with skills namespaced so an explicit call looks like /opencode-power-pack:code-review. Pi installs from the git URL with pi install git:github.com/waybarrios/opencode-power-pack, and pi list verifies what was discovered. Codex uses codex plugin marketplace add waybarrios/opencode-power-pack --ref main followed by codex plugin add opencode-power-pack@opencode-power-pack, after which the README says to start a new session and check /plugins. If you installed the sandbox CLI, the README's own first-use sequence is a doctor check followed by an execution under a named skill 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 doctor command reports on the isolation boundary from the active agent session, and the exec form runs a command with the profile the named skill is trusted with. If isolation is unavailable on your platform, the documented behaviour is to stop rather than run the command unsandboxed. ## Where the sandbox guarantee stops The fail-closed design is the right default, but it has a cost the README does not soften: on a platform where neither Seatbelt nor Bubblewrap is available, the sandbox path does not degrade to something weaker, it refuses. For a developer on an unsupported system, that means the 0.5.0 feature is simply unavailable, and the skills themselves are the whole product. The scope limit is the second constraint. Shell-contained means the command and everything it spawns stay inside the boundary. It does not mean the agent as a whole is confined, and the README states that automatic host routing and whole-agent isolation are not claimed yet. If your threat model is a compromised agent reading your filesystem, this runner is not that control, and the project is not pretending otherwise. There is also a judgment call in the profile design. Four profiles (observe, develop, network-read, publish) with explicit escalation checks assume the skill author classified each skill correctly. The README describes escalation as requiring confirmation, but the correctness of the initial classification is a property of the skill files, not of the runner. ## How this differs from Claude Code plugins and agent frameworks The closest comparison is a single-host plugin ecosystem. A Claude Code plugin marketplace entry installs skills for Claude Code and nothing else; the skill format is tied to that host's discovery rules. opencode-power-pack instead keeps one skills directory and writes three host manifests plus a pi entry around it, so the same content is reachable from four agents. The trade-off is that every host-specific quirk has to be handled in the packaging layer, which is why the repository carries .codexignore and a .plugin-scanner.toml at the top level. A different comparison is against agent frameworks that ship an execution runtime. Those give you a process model and an orchestration layer from the start. This project is content-first: fifty-four procedures plus, as of 0.5.0, a deliberately narrow command runner. If you want an agent platform, this is not one. If you want procedures that drop into the agent you already run, the packaging is the point. The repository also carries UPSTREAMS.json and THIRD_PARTY_NOTICES.md, which suggests some skills are adapted from other sources rather than written from scratch. The README does not detail that provenance in the excerpt available, so read those two files before assuming every skill is original work. ## Licence, upgrade cost and what to check first The npm package declares "license": "MIT AND Apache-2.0", while the repository's LICENSE file sits next to a LICENSES directory and a THIRD_PARTY_NOTICES.md. The AND is the part that matters: it is not a choice between two licences, and the per-component terms live in the notices file. That is consistent with a pack that adapts upstream material. Read THIRD_PARTY_NOTICES.md before redistributing anything from this repository in your own product; the README does not summarise which skills fall under which terms. Upgrade cost is low by design. Skills are copied files, and the documented update path re-copies them from a newer package with --force, which means local edits to a skill are overwritten. If you customise a skill, keep your version outside the install target or you will lose it on the next update. The global npm install is removed with npm uninstall --global @waybarrios/opencode-power-pack. Maintenance signals: the repository is not archived, and the last push was on 2026-09-07. Releases are frequent and close together, with v0.5.0 on 2026-08-17, v0.4.0 on 2026-08-13 and v0.3.0 on 2026-08-01. Fast minor releases also mean the skill set can shift under you between upgrades, which is another reason to pin a version if reproducibility matters to your team. ## Conclusion Adopt it if you already run Codex, OpenCode, Claude Code or Pi and want code-review and security-review workflows that arrive as files you can read and edit, rather than as prompts you retype each session. Skip it if you use none of those four agents, since every install path targets one of them, or if you need whole-agent isolation, because the README states the sandbox guarantee covers only the command and its descendants. Before installing, run the dry-run preview for the profile you intend to use, and check whether your Node version satisfies the >=20.11.0 engine requirement. ## FAQ ### How do I install opencode-power-pack? The npm route is npx @waybarrios/opencode-power-pack install --profile recommended, which copies skills into the shared .agents/skills location recognized by Codex, OpenCode and Pi. Claude Code, Codex and Pi each have their own install commands listed in the README. ### Is opencode-power-pack free to use? The repository is public and the npm package declares its licence as MIT AND Apache-2.0. Because the two licences are combined rather than offered as a choice, the per-component terms are in THIRD_PARTY_NOTICES.md and the LICENSES directory. ### Which coding agents does opencode-power-pack support? The README names four: Claude Code, Codex, OpenCode and Pi. The npm package requires Node >=20.11.0 and opencode >=1.18.7, and the skills are declared to Pi through the pi key in package.json pointing at ./skills. ### What does the sandbox in version 0.5.0 actually isolate? The README states the guarantee is that the command and its descendants are shell-contained, using Seatbelt on macOS and Bubblewrap on Linux. Automatic host routing and whole-agent isolation are explicitly not claimed. ### How do I update skills I already installed? The README shows npx @waybarrios/opencode-power-pack@latest install --profile review --force, which re-copies the skills from the latest package. Because it overwrites, any local edits to a skill file are replaced. ## Sources - [License: MIT](https://github.com/waybarrios/opencode-power-pack/blob/main/LICENSE) - [Project website](https://skills.waybarrios.com) - [README](https://github.com/waybarrios/opencode-power-pack/blob/main/README.md) - [Releases](https://github.com/waybarrios/opencode-power-pack/releases) - [waybarrios/opencode-power-pack on GitHub](https://github.com/waybarrios/opencode-power-pack) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/waybarrios-opencode-power-pack