plugin is the CyberOS plugin for Claude Code and Cowork - name cyberos, version 1.0.0, authored by CyberSkill. It packages the task workflow as host commands and a bundled skill, so a user can install CyberOS into a repo and drive its backlog without leaving the agent. tools/install/build.sh assembles it into dist/cyberos/, a plugin marketplace whose root carries .claude-plugin/marketplace.json and catalogs the plugin at plugin/ (its own manifest at plugin/.claude-plugin/plugin.json). The same build emits dist/cyberos/cyberos.plugin, a one-file bundle for hosts whose Add picker wants a file rather than a folder.

Install

• Claude Code: /plugin marketplace add <path to dist/cyberos>, then /plugin install cyberos@cyberos.
• Claude desktop and Cowork: add dist/cyberos/cyberos.plugin (the one-file bundle); the Add picker greys the Open button on a folder, so the file is the route here.

Both give the same commands and the ship-tasks skill.

The commands and the skill

• /install [repo] - install CyberOS into the current repo, or update it: autodetect the build / lint / test commands, vendor the machine under a gitignored .cyberos/, scaffold the docs/tasks/ backlog, and write the agent entry files. It prefers a full CyberOS payload when one is reachable and falls back to self-hosting from the plugin's own bundle. Idempotent - re-running it applies an update and never touches an existing BACKLOG.md, tasks, AGENTS.md, gates.env, or BRAIN.
• /update [repo] - compare the repo's installed version (.cyberos/VERSION) against an available payload and apply the update on request. Read-only until you confirm.
• /changelog [repo] - report the installed version and what changed recently.
• /help - what the plugin does: the commands, the task lifecycle, the two human gates, and where things live in an initialised repo.
• ship-tasks skill - drive the next eligible task from docs/tasks/BACKLOG.md through implement -> review -> test -> done. Type it as /ship-tasks, or just ask to ship a task and the skill is used. It bundles its own copy of the workflow doctrine (ship-tasks.md, EXECUTION-DISCIPLINE.md, STATUS-REFERENCE.md) so it works standalone.

Human-in-the-loop is required

The workflow halts at two human-acceptance gates: review acceptance (reviewing -> ready_to_test) and final acceptance (testing -> done). The agent runs everything it can verify between them, presents its review packet and test evidence, and stops for a recorded human verdict. It never sets done, and never pushes, merges, or deploys.

Doc-driven and agent-independent

The Claude plugin is convenience, not a dependency. The core is doc-driven: /install writes the canonical AGENTS.md spine, .cyberos/AGENT-ENTRY.md, and per-agent pointer files (all create-if-absent), and installs the ship-tasks skill natively into every skill-aware agent's folder (.claude/skills, .grok/skills, .commandcode/skills, .codex/skills, .opencode/skill). Point any file-and-shell agent - Codex, Cursor, Gemini, Grok, and the rest - at AGENTS.md or .cyberos/AGENT-ENTRY.md and it drives the same workflow, the same gates, the same required human verdicts. The per-agent instruction files, native skill directories, and MCP registration are catalogued in the agent-support matrix in tools/install/README.md.

One channel of many

dist/cyberos/ is delivered through several channels, of which the Claude plugin is one. The others include copying the folder, a git submodule or subtree, a curl | sh bootstrap, a GitHub Action that runs the machine gates in CI, a Docker image, a Makefile or just target, a Node stdio MCP server (tools task_install, task_gates, task_status, ship_task) for any MCP agent, an npx CLI (npx cs <command>), and a template-repo scaffolder (create.sh). The full catalog, with the trade-offs of each, is in tools/install/README.md; the install-and-operate walkthrough is in tools/install/docs/index.md.

Changelog

History lives in the changelog; this page describes only the current state.

Generated from modules/plugin/docs/index.md — edit the markdown source, not this file (TASK-DOCS-002).