codoop-flowcabin is an open-source Coding Agent plugin for creating small, offline Flow Cabin games from plain-language ideas.
The agent handles game design, visual direction, implementation, browser playtesting, and iteration. Deterministic Python scripts handle the work that must be exact: project placement, package validation, SHA-256 manifests, and ZIP creation.
you describe a game
│
▼
game design + technical art Sub-Agents
│
▼
editable HTML / CSS / JavaScript
│
▼
browser preview → validate → package
│
▼
flow-cabin-games/<game-id>/dist/<game-id>-<version>.zip
The plugin directory is treated as read-only. Generated games always live in the creator's coding workspace.
codex plugin marketplace add Codoop/codoop-flowcabin
codex plugin add codoop-flowcabin@codoop-flowcabinRestart or reopen Codex, choose /Codoop Flowcabin: Game, then say:
Make a one-button lantern game for Flow Cabin.
Alpha.5 changes the plugin identity from codoop-flow-cabin to match the
repository name. Remove the old identity before installing the new one:
codex plugin remove codoop-flow-cabin@codoop-flow-cabin
codex plugin marketplace remove codoop-flow-cabin
codex plugin marketplace add Codoop/codoop-flowcabin
codex plugin add codoop-flowcabin@codoop-flowcabin/plugin marketplace add Codoop/codoop-flowcabin
/plugin install codoop-flowcabin@codoop-flowcabin
Then invoke:
/codoop-flowcabin:game Make a small keyboard-controlled space game.
/plugin uninstall codoop-flow-cabin@codoop-flow-cabin
/plugin marketplace remove codoop-flow-cabin
/plugin marketplace add Codoop/codoop-flowcabin
/plugin install codoop-flowcabin@codoop-flowcabin
No SSH key? Add the marketplace with the full HTTPS URL:
https://github.com/Codoop/codoop-flowcabin.gitLocal development:
claude --plugin-dir /path/to/codoop-flowcabin
Prerequisite: Python 3.9 or newer. The generated games and packaging tools have no third-party runtime dependencies.
Describe the game, controls, and mood in one request:
/Codoop Flowcabin: Game Make a desktop game where I steer a moon rover left and right,
collect blue crystals, and avoid orange rocks.
The Skill will:
- Ask at most one question when a real gameplay choice is missing.
- Run game-design and technical-art reviews in parallel when Sub-Agents are available; level and narrative reviewers join only when needed.
- Create the game under the active coding workspace.
- Implement editable vanilla HTML, CSS, and JavaScript.
- Preview keyboard, pointer, pause, resume, restart, and resize behavior.
- Validate offline restrictions, paths, hashes, file count, and size limits.
- Produce an import-ready ZIP.
The remaining human step is to import the ZIP into Flow Cabin and confirm its cover, controls, pause, and resume behavior.
Every new game uses one predictable layout:
<workspace>/
└── flow-cabin-games/
└── <game-id>/
├── package/
│ ├── manifest.json
│ ├── index.html
│ ├── game.js
│ ├── styles.css
│ ├── cover.svg
│ └── assets/
└── dist/
└── <game-id>-<version>.zip
package/is the editable runtime source and the only content placed in the ZIP.dist/contains generated release files.- Existing games keep the same
manifest.json.idand receive an explicit version update. - Another output root is used only when the creator explicitly requests it.
- The initializer refuses to overwrite an existing game or write into the installed Skill directory.
See project-layout.md for the
complete output contract.
skills/game/SKILL.md defines the creation,
playtest, review, validation, and delivery sequence.
Review-only roles live in skills/_shared/:
game-designer— goal, core loop, controls, feedback, and retry behavior.technical-artist— original visual direction, hierarchy, HUD, and cover.level-designer— maps, waves, stages, puzzles, and difficulty progression.narrative-designer— premise, dialogue, missions, and choices.
The main agent owns all file edits. Experts return bounded decisions, risks, and acceptance checks, so parallel reviews cannot overwrite the game.
create_game.pycreates the standard project in the user's workspace.flow_cabin_package.pyvalidates and packages the runtime using only the Python standard library.
The Skill ships a playable vanilla starter plus frozen references for project layout, package security, runtime lifecycle, and expert orchestration.
| Guarantee | Behavior |
|---|---|
| Offline-first | Rejects remote URLs, network APIs, iframes, dynamic imports, and Node/Electron APIs |
| Safe paths | Rejects traversal, hidden package paths, symbolic links, and missing local resources |
| Exact manifest | Rewrites and verifies SHA-256 for every runtime file except manifest.json |
| Size limits | Maximum 1,000 files, 100 MiB unpacked, and 25 MiB ZIP |
| Native runtime | Uses local HTML, CSS, JavaScript, SVG, PNG, WebP, and JSON only |
| Browser fallback | Runs without window.FlowCabinGame; subscribes to pause/resume when available |
Only package/ enters the archive, and the ZIP root contains
manifest.json and index.html directly.
Coding Agents call these scripts automatically. For local development:
python3 skills/game/scripts/create_game.py /path/to/workspace my-game \
--title "My Game"
python3 skills/game/scripts/flow_cabin_package.py pack \
/path/to/workspace/flow-cabin-games/my-game/package \
--output /path/to/workspace/flow-cabin-games/my-game/dist/my-game-1.0.0.zip
python3 skills/game/scripts/flow_cabin_package.py validate \
/path/to/workspace/flow-cabin-games/my-game/packagecodoop-flowcabin/
├── .agents/plugins/marketplace.json
├── .claude-plugin/
│ ├── marketplace.json
│ └── plugin.json
├── .codex-plugin/plugin.json
├── skills/
│ ├── _shared/ # expert review roles
│ └── game/
│ ├── SKILL.md # agent orchestration
│ ├── references/ # output, runtime, package, and review contracts
│ ├── scripts/ # deterministic initializer and packager
│ └── templates/ # editable vanilla starter
├── tests/
├── docs/install.md
├── CHANGELOG.md
├── README.md
└── README.zh-CN.md
English documents use suffix-less filenames. Chinese documents use the
.zh-CN suffix.
python3 -m unittest discover -s tests -vThe tests run without AI or network access. They cover project initialization, standard output placement, ZIP layout, hashes, offline restrictions, traversal, symbolic links, bad entry points, missing resources, and package limits.
| Agent | Status | Installation |
|---|---|---|
| Codex Desktop | First-class | Codex plugin marketplace |
| Codex CLI | First-class | Codex plugin marketplace |
| Claude Code | First-class | Claude plugin marketplace |
| Cursor / Gemini | Generic | Copy the Skill and let the agent read SKILL.md |
If the host does not support Sub-Agents, the main agent applies the same expert roles sequentially.
Where are generated games stored?
Under <workspace>/flow-cabin-games/<game-id>/ by default, never inside the
installed plugin.
Can I choose another output directory? Yes. Ask for it explicitly. Otherwise the standard workspace path is used.
Can the Skill update an existing game?
Yes. It edits the existing project in place, preserves the game ID, increments
the version, and replaces the ZIP in that project's dist/.
Does this use React, Phaser, Vite, or npm? No. Games use editable vanilla HTML, CSS, and JavaScript.
Does a passing validation mean it was tested inside Flow Cabin? No. Browser and package validation are automated; real Flow Cabin import is the creator's final acceptance step.
Why are source and ZIP separated?
The source remains easy to edit, while dist/ is safe to delete and regenerate.
docs/install.md— local installation and generated project behavior.CHANGELOG.md— release history.skills/game/references/project-layout.md— generated output contract.skills/game/references/package-v1.md— package and manifest contract.skills/game/references/runtime-api.md— Flow Cabin lifecycle integration.