dstack

hey, i'm hungrykelvin123. i've been developing games for about two years now, and ever since ai became much more prominent in the community, i've wanted a faster, deeper, and more token-efficient way to write, debug, and ship code for roblox games.

that's when i found pstack, a cursor plugin with a bunch of skills designed by a highly skilled engineer to help agents write high-quality code. unfortunately, i don't use cursor, so i used pstack ported to codex instead.

this worked pretty well when i was designing the systems of my game, but one thing always nagged at me. because pstack was a generalized engineering plugin and wasn't made specifically for roblox, i found that my agents did much more than they needed to and wasted tokens. when it comes to verifying code inside roblox, i also absolutely despise when ai uses the roblox studio mcp server to playtest for me. it wastes a ton of tokens on minor fixes that i probably would've found myself.

that's why i created my own version, dstack. dstack is specifically made for roblox and expects you to have the roblox studio mcp server available. i do all of my scripting in visual studio code and sync it to my game with rojo, but i keep the mcp server open so my agent can inspect instances when the files alone aren't enough. dstack is most accurate and produces its highest-quality work when mcp is connected to an open roblox studio session. most importantly, dstack explicitly forbids the agent from using the mcp server to playtest the game, because that only produced useless results for me.

have fun coding!

what dstack is

dstack is a plugin for Claude Code and Codex. it is not a model or a hosted service. it supplies one shared Roblox/Luau skill tree, focused playbooks, optional read-only reviewer agents, client-specific manifests, and guarded hooks.

the repository uses the same distribution shape as mature multi-client plugins:

.
├── .agents/plugins/marketplace.json       # Codex marketplace
├── .claude-plugin/marketplace.json        # Claude Code marketplace
└── plugins/dstack/                        # the canonical plugin
    ├── .claude-plugin/plugin.json
    ├── .codex-plugin/plugin.json
    ├── .codex-plugin/prompts/             # explicit Codex prompt shims
    ├── agents/                            # Claude read-only reviewer lanes
    ├── hooks/                             # separate Claude and Codex hook configs
    ├── models.json                        # bounded reviewer recommendations
    ├── references/                        # shared Roblox contracts
    └── skills/                            # one shared skill tree

the two clients load the same plugins/dstack/skills files. only the manifest, prompt surface, hook schema, and subagent dispatch differ. client caches are implementation details: a path under plugins/cache is a versioned snapshot, not a stable skill path to paste into a prompt.

install

install the marketplace that matches the client you use. you do not need both.

Claude Code

Run these commands in Claude Code:

/plugin marketplace add HungryKelvin123/dstack
/plugin install dstack@dstack
/reload-plugins

Claude exposes the skills as /dstack:<skill>. If you are developing from a checkout, test it without installing with:

claude --plugin-dir ./plugins/dstack

Codex

Run these commands in a shell:

codex plugin marketplace add HungryKelvin123/dstack --ref main
codex plugin add dstack@dstack

Codex exposes the same skills by name. Start a new Codex task after installation or an update so the skill catalog is rebuilt. If you previously installed the old dstack-local marketplace, remove that entry and install the dstack entry once; do not keep two copies enabled.

local validation

Before publishing a checkout, validate the two manifests and the Claude marketplace with the client tools available on your machine:

claude plugin validate ./plugins/dstack
codex plugin list --json

If a client is not installed, use the repository tests below; do not claim that client runtime loading was verified.

setup

After installation, run the explicit setup skill once when you want client routing or reviewer lanes checked:

Claude Code:

/dstack:setup-dstack

Codex:

Use $setup-dstack to configure DStack for this client.

setup verifies the active client, the canonical skill tree, the automatic unslop rule, the Roblox Studio MCP state, and the optional reviewer panel. It does not edit your game. It does not write global instructions unless you approve a single bounded startup block. If MCP is missing or unconfigured, it follows the stop-and-ask flow in the MCP setup reference; after configuration it ends the turn and directs you to restart the client.

get started

use David Mode for a non-trivial Roblox task:

Claude Code:

/dstack:david-mode Trace why this remote can award a reward twice, fix the root cause, and give me the exact Studio checks to run.

Codex:

$david-mode Trace why this remote can award a reward twice, fix the root cause, and give me the exact Studio checks to run.

David Mode reads repository instructions and source first, chooses the smallest Roblox playbook, routes to focused skills only when needed, runs local checks such as rojo build, and reports what you must test in Studio. For a narrow request, invoke one focused skill directly instead of paying for the full router. unslop is the only automatically invoked skill; all other dstack skills are explicit.

skills

skill	use it for
david-mode	route a non-trivial Roblox task through a bounded workflow.
setup-dstack	configure and verify the client integration.
how	trace a Roblox subsystem through modules, instances, remotes, replication, and persistence.
why	recover why an implementation exists from code, history, and decisions.
recall	rebuild recent project context before resuming work.
architect	design ownership, typed Luau modules, remotes, and persistence boundaries.
blast-radius	find what a change could break across the Roblox project.
luau-best-practices	review or write maintainable typed Luau.
roblox-security	review authority, remotes, purchases, rewards, RNG, persistence, and abuse cases.
roblox-performance	diagnose client, server, physics, rendering, replication, or memory performance.
roblox-data	design safe persistence, schemas, session ownership, retries, and cross-server state.
roblox-networking	design remote protocols, replication, validation, and reliability.
roblox-monetization	build retry-safe developer product, pass, and subscription flows.
roblox-ui	build responsive touch, keyboard, mouse, and gamepad interfaces.
roblox-physics	design assemblies, constraints, collision, hit detection, and network ownership.
arena	compare independent attempts at the same bounded problem.
swarm	split independent implementation slices when the user requests parallel work.
interrogate	run a risk-budgeted adversarial review with independent read-only model lanes.
tdd	write a cheap deterministic regression test before a fix.
no-comments	remove narrating comments while preserving real constraints.
create-verification-skill	create a reusable verification skill for a Roblox repository.
maintain-verification-skill	update a verification skill when the project changes.
show-me-your-work	keep a reviewable decision log during long work.
figure-it-out	design a bounded playbook when none of the bundled ones fit.
teach	explain a system using implementation and history.
reflect	capture lessons from a completed task in the skills themselves.
automate-me	turn a recurring working style into a personal mode skill.
technical-writing	write or review technical documentation.
unslop	remove AI tells and vague prose; this one runs automatically.
bro	restate a technical answer in plain language.

David Mode also reads the small principle-* skills only when their rule fits the task. They are intentionally not user-facing defaults; see the complete list under plugins/dstack/skills.

Interrogate and model use

The old Interrogate wording implied a configured panel but did not guarantee that the client actually served different models. The current skill makes that distinction explicit:

• one reviewer for a local change;
• two for a cross-module change;
• three for security, saved data, monetization, consequential RNG, networking, or lifecycle boundaries;
• four only for a critical or explicitly maximum review.

Claude uses the bundled read-only haiku, sonnet, and opus reviewer agents when those model aliases are available. Codex requests distinct advertised model overrides (gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna) when the active host supports them. Other clients use only observable configured lanes. A missing, failed, or unverified lane is reported; it is never silently replaced or counted as model diversity. All reviewers receive the same evidence and rubric, and Interrogate never applies their findings automatically. See models.json and the panel contract.

Every model run costs its own allowance. Use Interrogate when independent challenge is worth that cost; use blast-radius or a focused skill for a routine change. Arena and Swarm remain explicit and are not called by setup automatically.

Roblox Studio MCP boundary

dstack is most accurate when the Roblox Studio MCP server is open and connected to the target Studio session. It does not bundle or auto-start that server; configure it for your client using the official Roblox Studio MCP guide.

With a usable local repository and Rojo/project sync, the repository remains authoritative and MCP supplies missing instance context. If Rojo/project sync is unavailable, or no local repository is both present and connected to the target Studio, dstack may use all non-playtest MCP operations the server exposes within the user's requested scope: inspect and edit scripts, instances, properties, hierarchy, attributes, tags, attachments, and authored assets when supported. It inspects before mutating, keeps writes narrow, and reports exact Studio paths. Those edits are not repository diffs, tests, or Rojo proof.

If the MCP server is configured but Studio is closed or disconnected, dstack tells you once and continues from repository evidence when possible. If the runtime explicitly reports the server is missing or unconfigured, dstack stops and asks whether you want setup. yes follows the official guide and then ends the turn so you can restart Claude Code/Codex. no continues without MCP and includes one concise notice that MCP gives DStack its fullest Roblox context and accuracy. No connected Studio alone is never treated as proof of an uninstall.

dstack does not use Roblox Studio MCP to playtest. It never launches, starts, stops, simulates, or controls a test session. The user performs all Studio playtesting. After local checks pass, dstack returns a short manual Studio checklist for you.

update or remove

Claude Code:

/plugin marketplace update dstack
/plugin update dstack@dstack

Codex:

codex plugin marketplace upgrade dstack
codex plugin remove dstack@dstack
codex plugin add dstack@dstack

Start a new task after an update. If an old dstack-local entry remains, remove it from the client before reinstalling dstack; versioned cache directories are normally cleaned by the client and should not be edited by hand.

To remove the plugin:

/plugin uninstall dstack@dstack

codex plugin remove dstack@dstack
codex plugin marketplace remove dstack

develop and verify

Run the repository checks from the root:

node --test tests/*.test.mjs

The tests verify the nested marketplace structure, both manifests, prompt shims, internal links, invocation policy, David Mode state, the missing-versus-closed MCP setup gate, the reviewer model-diversity contract, and the permanent Studio no-playtest boundary.

origin and license

dstack is a Roblox-focused derivative of pstack by Lauren Tan, adapted from pstack-for-codex by Aqua-123. it is not an official Cursor, Roblox, Anthropic, or OpenAI project.

MIT. See LICENSE and NOTICE.