MOStudio Kit

Agent skills for operating live game projects — not writing new ones.

MOStudio Kit is distributed from the GameStudio-CodexKIT repository. Repository URLs, install commands, and the stable plugin ID game-studio-codex-kit intentionally keep that technical name.

Most gamedev AI skills teach an agent to build games. This kit teaches it to keep a shipped game alive: crash triage on a C++ MMORPG server, offline-mode debugging in a legacy Unity client, MySQL migrations that can't eat player saves, Lua client/server contract audits, liveops incidents, store submissions — with an evidence system that makes it structurally hard for the agent to lie about what it did.

you:    "Client can't enter the game offline. Fix it."

agent:  → routes to unity-client-offline-debugging (read-only diagnosis first)
        → traces offline flag through bootstrap, finds stale generated Lua mirror
        → Verified: regenerated via owner pipeline, exit 0, artifact path attached
        → BLOCKED: play-mode re-audit — editor lost mid-run. Not faked as PASS.

That last line is the point. Every claim carries one of four labels — Verified / Snapshot / Unverified / BLOCKED — and an unavailable runner is BLOCKED, never PASS. Your agent's report becomes trustworthy precisely because it is allowed to say "I couldn't verify this."

Visual tour

The kit connects a live game, its studio stack, and its operating evidence into one routed workflow. No dashboard theater: every page below maps to a real contract in the catalog.

Built for live production Clients, services, generated assets, and player data stay inside the risk model.	Ask for the outcome The router selects the smallest matching specialist workflow.
One studio catalog 47 canonical skills, seven packs, six thin persona lenses, and 22 canonical agent roles.	Trust is explicit state BLOCKED is never converted into PASS.
Mutation gates scale with risk Diffs, restore paths, reviewers, dry runs, and human approval stay visible.	Dogfooded on a live MMORPG Real fixes shipped; missing proof stayed honestly blocked.

Install paths — Codex App/CLI and Hermes Agent

Why this kit is different

	Typical skill packs	MOStudio Kit
Target	Writing new code	Operating a live, fragile, revenue-bearing game
Failure mode	Agent claims success	BLOCKED verdicts + PASS requires command, exit code, artifact
Mutations	Agent edits freely	4-tier risk gates: read-only → low → medium (reviewer+backup) → high (human approval + dry-run)
Multi-agent	Single session	One-writer-per-file ownership, lock protocol, handoff contracts, review/bug-hunt swarms
Routing quality	Hope the description matches	276 deterministic eval cases incl. negative + collision cases, bilingual (EN/VI)

Field evidence — real run on a live MMORPG

The kit's operating model has been dogfooded on a live Unity 6 WebGL MMORPG client (Lua gameplay layer, workbench localization pipeline, ~10 concurrent agent-session lock namespaces). A skill-routed Codex run reviewed and fixed mixed-Chinese player-facing text with evidence labels on every claim:

Gate	Result
Config-gap batch (32 IDs)	PASS with hash-verified artifacts; catalog 25,609 → 25,641
Sanitizer RED test	GREEN (10/10 focused, 34/34 combined)
Overflow string	Fixed via signature-preserving %.0s, edit-mode verified (150 px in 160 px)
Strict doctor	BLOCKED (stale artifact outside approved lock scope) — not faked
Play-mode re-audit	BLOCKED (editor instance lost mid-run) — not faked
Collateral damage	None: 0 foreign locks touched, 0 commits, unrelated work preserved

Two runs, three real fixes shipped through owning pipelines, two blockers left standing because the evidence to close them did not exist. The full narrative, including real runner failures and the human approval gate, is in docs/case-studies/unity-mmorpg-global-localization.md.

Quick install

Choose one primary runtime. A repository clone is not required for normal use.

Runtime	Install
Codex App/CLI	codex plugin marketplace add hoatv2211/GameStudio-CodexKIT, then install MOStudio Kit from /plugins or the App Plugins UI
Hermes Agent	npx skills add hoatv2211/GameStudio-CodexKIT -a hermes-agent -g -y

Start a new Codex or Hermes session after installation, then ask for the outcome directly. The root router selects the smallest matching studio workflow.

What you can ask

No slash commands, no skill names required — describe the problem and the router picks the workflow:

You say	Routed skill
"Audit this game project and route the work to the right studio skills"	studio-project-intake (start here on any new repo)
"Server crashed, here's the stack trace"	cpp-server-crash-triage
"Client can't enter offline mode"	unity-client-offline-debugging
"This UI item doesn't render / wrong draw order"	unity-ui-rendering-debugging
"I need to change the items table schema"	game-database-migration-safety (dry-run gated)
"Is this Lua field actually validated server-side?"	lua-client-server-contract-audit
"Players found a dupe exploit, live incident"	liveops-incident-response
"Ship-readiness check before the store build"	release-candidate-preflight
"Write a handoff so the next session can continue"	studio-handoff

Add "evidence labels mandatory" to any request when you want the strict verification contract enforced end-to-end.

Install in Codex

Add the GitHub marketplace from a terminal:

codex plugin marketplace add hoatv2211/GameStudio-CodexKIT
codex

In the Codex CLI session, enter /plugins, select the MOStudio Kit marketplace, open MOStudio Kit, and install it. Start a new CLI session before using the bundled skills.

In the ChatGPT desktop app, open Plugins from Codex, select the MOStudio Kit marketplace, install the plugin, and start a new task. Ask for the outcome directly or invoke the plugin with @game-studio-codex-kit when you want an explicit route.

Example first request:

Audit this game project and route the work to the right studio skills.

Update the marketplace snapshot, then review the plugin in /plugins and start a new session:

codex plugin marketplace upgrade gamestudio-codex-kit

To remove it, uninstall the plugin from /plugins first, then remove the marketplace:

codex plugin marketplace remove gamestudio-codex-kit

The repository root is the plugin package. .codex-plugin/plugin.json exposes the canonical skills/ directory, while .claude-plugin/marketplace.json is the Codex-supported compatibility marketplace discovered from the GitHub repository. Each domain skill carries its own generated helper scripts, so the installed plugin does not depend on the user's current working directory.

Install in Hermes Agent

Install all 47 skills globally with the Agent Skills CLI:

npx skills add hoatv2211/GameStudio-CodexKIT -a hermes-agent -g -y

Start a new Hermes Agent session after installation. The installer discovers the repository-root skills/ catalog and copies each skill directory into the Hermes global skills directory. Bundled helpers therefore live inside the skill that uses them. Omit -g when you intentionally want a project-local .hermes/skills/ installation.

To inspect the catalog before installing:

npx skills add hoatv2211/GameStudio-CodexKIT -a hermes-agent -l

Repository governance tools such as the catalog validator, originality audit, and adapter generators are intentionally not copied by Hermes. Clone the repository only when maintaining the kit itself.

Current state

• 47 canonical skills, including one root entry router and 46 routed workflow or domain skills.
• 276/276 deterministic Tier-A routing cases.
• 10 external-catalog collision cases against six generic neighboring skills.
• 12 governed real-project dogfood scenarios defined with strict PASS/BLOCKED evidence validation.
• Seven installable packs, six thin personas, and 22 canonical agent roles.
• Two primary distributions: native Codex plugin installation and Agent Skills CLI installation for Hermes Agent.
• Eighteen standalone skills with 21 generated helper copies and explicit full-clone boundaries for repository-only governance tools.
• Optional generated exports for manual Hermes, Codex, pack, project-local .agents/ skills, and .codex/agents/ role overlays.
• Structural, provenance, secret, network/package, safety, external-catalog collision, behavior, pressure, and lifecycle gates.
• The 47-skill catalog is beta as a studio-adopted kit based on maintainer-confirmed use in the FPC project (Snapshot); this does not claim that every skill ran individually. Deterministic gates and the localization case study are Verified, while per-skill Tier-B, behavior, pressure, and runtime evidence remain required for stable.

The template does not claim that a Unity build, C++ server, database migration, store submission, or liveops action has run unless a real project artifact proves it.

Catalog

Pack	Focus
studio-core	Intake, multi-repository routing, agent orchestration, design, planning, debugging, evidence, safe mutation, review, handoff, and skill governance
unity	Offline client debugging, UI rendering, localization, GUID/meta integrity, and batch builds
cpp-lua-mmorpg	Local services, MySQL safety, Lua contracts, C++ crashes, packet protocols, authority, and save migration
production-design-liveops	Playtests, performance, economy, balance, release, stores, incidents, and telemetry

The authoritative capability list is registry/capabilities.yaml. Pack composition and persona routes live in registry/packs.yaml and registry/personas.yaml.

Maintain this repository

A full repository clone exposes an internal maintenance bundle that is intentionally separate from the 47 distributed skills and 22 cataloged agent roles:

• Invoke codexkit-repository-maintenance for CI, governance, catalog, generator, adapter, packaging, documentation, architecture, versioning, or release-readiness work on this repository.
• Select the repository-local codexkit-maintainer agent when the task needs a bounded writer and integration owner.
• The skill verifies .codex-plugin/plugin.json, registry/capabilities.yaml, scripts/validate.py, and skills/; another repository receives BLOCKED: repository identity mismatch.
• Internal files live under .agents/skills/ and .codex/agents/. They are excluded from registries, packs, generated adapters, project scaffold templates, and installed game projects.
• Follow workflows/repository-maintenance.md and run the local gates below before handoff.

Contributor requirements

• Python 3.11+
• PyYAML
• jsonschema
• Git

Codex users need a supported Codex surface with plugin marketplace support and Git. Hermes Agent users need Node.js/npm for the npx skills installer. Bundled deterministic helpers use Python 3.11+ and may require PyYAML when they validate or render YAML-backed project profiles. A full clone requires PyYAML for repository validation, routing, policy, profile, and generation tools, plus jsonschema for governed dogfood result validation. The project uses standard-library unittest; pytest is not required.

Verify the kit

Run the deterministic local gates before distributing or handing off changes:

python -B scripts/sync_skill_resources.py . --check
python -B -m unittest discover -s tests -p "test_*.py"
python -B scripts/validate.py .
python -B scripts/route_eval.py .
python -B scripts/secret_scan.py .
python -B scripts/policy_check.py .
python -B scripts/external_collision_eval.py .
python -B scripts/doctor.py --check --root .

To compare against an installed Codex/Hermes catalog as well as the checked-in generic snapshot, repeat --external-root for each catalog root:

python -B scripts/external_collision_eval.py . --external-root /path/to/installed/skills

Install the managed pre-commit hook:

python -B scripts/doctor.py --install-hook --root .

The hook runs structural validation, Tier-A routing, and secret scanning. It never commits or pushes. On Linux and macOS, hook installation also marks .git/hooks/pre-commit executable; Windows does not expose POSIX execute bits.

Use in a game project

The commands in this section are maintainer workflows and require a full repository clone. Normal Codex and Hermes installations are prompt-driven and use their installed skill catalogs directly.

Inspect a new project without writing files:

python -B scripts/project_scaffold.py D:/path/to/game-project

Applying the scaffold is a medium-risk mutation and requires a reviewer plus backup root:

python -B scripts/project_scaffold.py D:/path/to/game-project --apply --reviewer "QA Lead" --backup-root D:/path/to/game-project/.scaffold-backup

The per-project adapter is report-only by default. Capture its proposed plan first; this command creates nothing:

$report = python -B scripts/generate_adapters.py . --target per-project --output D:/path/to/game-project | ConvertFrom-Json
$report | ConvertTo-Json -Depth 10

Review the full report and its plan_digest, then apply that exact plan. Apply requires a named reviewer, a disjoint project-local backup root, and the approved plan digest:

python -B scripts/generate_adapters.py . --target per-project --output D:/path/to/game-project --apply --reviewer "QA Lead" --backup-root D:/path/to/game-project/.adapter-backup --plan-digest $report.plan_digest

Apply rebuilds the plan and refuses mutation if the source catalog, project profile, or any managed target changed after the report. If it refuses, generate and review a new report instead of reusing the old digest. The backup root must not overlap any planned target such as .agents/ or .codex/.

The apply command writes kit-owned skills under .agents/skills/ and builds a role overlay from packaged generic agent templates plus the profile specialist overlay under .codex/agents/. It emits an inert activation file at .codex/agents.generated.toml, records per-file ownership hashes in .agents/registry.json, and returns a safe-mutation manifest plus restore command. The activation file has no effect until it is manually reviewed and merged into project configuration. The adapter leaves .codex/config.toml untouched and never overwrites .codex/config.toml.

Uninstall is hash-safe: it removes only files whose current content still matches their recorded per-file ownership hash, preserves unmanaged or drifted local agents, and reports PARTIAL recovery with the remaining owned paths when safe removal cannot finish. Resolve those paths from the report rather than deleting project-local content blindly.

Advanced maintenance: packs and adapters

Build all packs:

python -B scripts/build_packs.py . --output dist/packs

Generate standard adapter exports locally from canonical skills when maintaining a manual distribution:

python -B scripts/generate_adapters.py . --target hermes --output adapters/hermes
python -B scripts/generate_adapters.py . --target codex --output adapters/codex

Native Codex and Hermes Agent installation both consume canonical skills/ and do not require adapter trees. adapters/ is ignored generated output; its files and generated packs carry a Do not edit manually marker. Change skills/ or registry/, then regenerate any export artifacts you intentionally distribute outside this repository.

Bundled skill helpers are also generated. Change the canonical source under root scripts/, update registry/skill-resources.yaml, and run:

python -B scripts/sync_skill_resources.py .

Governed evaluation

Tier-A routing is deterministic. Behavior, pressure, and Tier-B routing require result artifacts from a governed runner; skill wording alone is never evidence of correct model behavior.

Export cases and local BLOCKED status into the ignored evidence/local/ workspace:

python -B scripts/behavior_eval.py . --export evidence/local/behavior-cases.jsonl
python -B scripts/behavior_eval.py . --status evidence/local/behavior-status.json
python -B scripts/pressure_eval.py . --export evidence/local/pressure-cases.jsonl
python -B scripts/pressure_eval.py . --status evidence/local/pressure-status.json
python -B scripts/tier_b_eval.py . --export evidence/local/tier-b-cases.jsonl
python -B scripts/tier_b_eval.py . --status evidence/local/tier-b-status.json
python -B scripts/dogfood_eval.py . --export evidence/local/dogfood-cases.jsonl
python -B scripts/dogfood_eval.py . --status evidence/local/dogfood-status.json

The dogfood pack contains twelve real-project scenarios covering Unity offline/bootstrap and NGUI rendering, batchmode builds, MySQL safety, local service ports, C++ crashes, Lua contracts, liveops incidents, release preflight, project intake, workspace routing, and agent orchestration. Supply governed results with --results; missing Hermes/live-project execution remains BLOCKED.

After a governed runner produces the strict {"results": [...]} object and stores hashed artifacts under an approved root, validate it and generate promotion-eligible summaries:

python -B scripts/dogfood_eval.py . --results evidence/local/dogfood-results.json --artifact-root evidence/local --summary-dir evidence/local/dogfood

The evaluator rejects bare PASS labels, schema drift, unsafe or missing artifact paths, digest mismatches, non-zero exits, missing reviewers or project snapshots, unauthorized writes, and incomplete case coverage. Legacy array results remain diagnostic-only and cannot generate promotion summaries. See docs/authoring/dogfood.md for the runner contract.

When a runner, live project, engine, service, or permission is unavailable, the correct verdict is BLOCKED, never a fabricated PASS.

scripts/check_originality.py is also BLOCKED until the upstream snapshots in registry/upstream-sources.yaml are restored or supplied explicitly.

Evidence labels

• Verified: observed command, test, build, primary source, or artifact.
• Snapshot: true for a named commit, version, configuration, or environment.
• Unverified: hypothesis or forecast without sufficient evidence.
• BLOCKED: required runner, project, dependency, permission, or tool is unavailable.

A PASS claim includes the command, exit code, and artifact path when applicable.

Repository layout

 .codex-plugin/ native Codex plugin manifest
 .claude-plugin/ GitHub marketplace metadata read by Codex
skills/       canonical workflows plus generated self-contained helper copies
registry/     capability, pack, persona, skill-resource, and upstream indexes
personas/     thin role lenses and routes
workflows/    human-facing entry workflows
scripts/      deterministic helpers, generators, and gates
tests/        unittest regression and governance coverage
evals/        routing, external-catalog, dogfood, behavior, pressure, and schema fixtures
adapters/     ignored local Hermes and Codex export output
evidence/     reusable example; local runner output is ignored
.archive/     ignored local planning history, never distributed or active policy

Architecture details live in docs/architecture/overview.md; skill authoring rules live in docs/authoring/skills.md; operating rules live in AGENTS.md.

Local archive boundary

.archive/ is ignored local history for completed plans or cleanup notes that an individual maintainer wants to retain. It is not part of the public template, plugin, adapters, or packs, and agents must never treat it as current instructions. Durable project decisions belong in maintained docs; reproducible runtime evidence belongs in evidence/local/ or CI artifacts.

Contributing skills

Use update-first authoring:

1. Search the registry for overlap.
2. Add a failing routing, behavior, pressure, or deterministic regression case.
3. Update the smallest canonical skill and provenance record.
4. Update registry/skill-resources.yaml and regenerate bundled helpers when a skill uses a root helper.
5. Run all local gates.
6. Validate the native plugin and regenerate any packs or adapters being distributed.
7. Keep new skills draft or experimental until studio adoption is confirmed; use beta for adopted workflows, and require governed promotion evidence for stable or release.

License

MIT. Third-party influence is recorded through per-skill provenance. CC BY-NC-SA sources are pattern-only and must not be copied into the kit.