Hosting the AI Slop Party

House Rules

The file set from The AI Slop Party. Rooms by blast radius, gates at the door, and agents that read the fridge.

Source on GitHub

Running the house outside Claude Code

The rules, gates and agents here are written once and read by more than one runtime. This page says what each runtime reads, what we generate for it, and what does not carry over. Checked against official docs on 2026-10-04; links at the bottom. Anything marked unconfirmed was not in an official page.

Contents

What is portable as-is

File Read natively by
AGENTS.md Codex, GitHub Copilot, Cursor; Gemini CLI with "context": {"fileName": ["AGENTS.md", "GEMINI.md"]} in .gemini/settings.json
platform/AGENTS.<stack>.md every runtime, because AGENTS.md tells the agent which one to open for the files in the diff. Copilot and Cursor also get them auto-attached by path (generated, below)
.github/ gates, scripts/, adr/ runtime-independent
.claude/agents/*.md, .claude/skills/ Claude Code, and also Copilot and Cursor, which scan those paths
.claude/agent-memory/ plain markdown; Claude Code loads it as agent memory, any other agent reads MEMORY.md on instruction

What we generate

scripts/export-agents.py reads agents/**/*.md and skills/*/SKILL.md (the only sources you edit) and writes the layouts below. --check fails if a generated copy is stale.

flowchart LR
  A["agents/**/*.md<br/>skills/*/SKILL.md"] --> X[scripts/export-agents.py]
  X --> S[".agents/skills/<br/>Codex · Copilot · Cursor · Gemini"]
  X --> C[".codex/agents/*.toml"]
  X --> G[".gemini/agents/*.md"]
  X --> P[".github/agents/*.agent.md<br/>.github/instructions/*.instructions.md"]
  X --> R[".cursor/rules/*.mdc"]
  A --> SDK["agents/sdk/openai_agents_house.py<br/>OpenAI Agents SDK loader (reads at runtime)"]
Output For What changes from the source
.agents/skills/<name>/SKILL.md Codex, Copilot, Cursor, Gemini frontmatter reduced to the Agent Skills spec fields (name, description, license, compatibility, metadata, allowed-tools)
.codex/agents/<name>.toml Codex name, description, developer_instructions = the body
.gemini/agents/<name>.md Gemini CLI name, description, kind: local, model: inherit; no tools so it inherits all
.github/agents/<name>.agent.md Copilot name, description; bodies over 30,000 characters are trimmed to the intro plus a pointer to the full file
.github/instructions/<stack>.instructions.md Copilot the platform file with applyTo globs
.cursor/rules/<stack>.mdc Cursor the platform file with globs

The path globs live in PLATFORM_GLOBS at the top of the script. They assume the AGENTS.md layout; change them with it.

Per runtime

OpenAI Codex (CLI, IDE, cloud)

OpenAI Agents SDK (Python; JS has the same shape)

agents/sdk/openai_agents_house.py builds Agent(name, instructions, handoff_description, tools) for every definition. host-agent and eng-manager-agent get the others via agent.as_tool(), so the manager keeps control and synthesizes, which is how they are written. You supply function tools per capability (read, write, shell, web); the Claude tool names in the markdown map onto those. Nothing in this repo depends on the SDK; install openai-agents in your project.

from openai_agents_house import load_house
from agents import Runner
house = load_house(tools={"read": [read_file], "shell": [run_shell]})
print(Runner.run_sync(house["host-agent"], "Review PR 123 against our architecture.").final_output)

GitHub Copilot (VS Code, coding agent)

Cursor

Gemini CLI

What does not carry over

Claude Code feature Elsewhere
tools: allowlists (Read, Grep, Bash, …) ignored; Copilot and Gemini default to all tools, Cursor always inherits. The SDK loader maps them to capabilities
memory: project auto-loading the folder is plain markdown; the persona files tell the agent to read MEMORY.md first
Skill frontmatter context: fork, agent:, background:, argument-hint: stripped in .agents/skills/; the skill body says how to run without them
$ARGUMENTS substitution the skill body also accepts the subject from the request text
claude --agent host-agent load agents/person/host-agent.md (or the exported TOML / .agent.md) as the session’s instructions
smart-third-grader’s ~/.claude/.third-grader-on toggle and hook the voice rules work anywhere; the toggle is Claude Code’s hook mechanism

Keep the copies honest: scripts/export-agents.py --check before a PR that touches agents/ or skills/. Those paths, and every generated directory, are Kitchen.

Sources