codex-subculture

I use Codex Subculture as a Sol bridge between me and Luna Max. I stay in a normal Sol thread. When I activate the protocol, Sol inspects the project, decides the architecture, creates a new, first-class Luna Max thread, and gives it a complete implementation contract. Sol then ends its turn. Luna signals the Sol thread only when blocked or ready for verification, and Sol returns to inspect and integrate the result.

This is thread orchestration, not a native subagent loop. The Luna worker has its own threadId, visible history, and separate entry in the Codex task list.

The protocol activates only when the user gives the exact standalone instruction SUBCULTURE. Mentioning, reviewing, or translating the word does not activate it.

How the bridge works

1. I describe the task in a normal Sol thread and activate SUBCULTURE.
2. Sol inspects the project and owns the architecture, plan, and material technical decisions.
3. Sol creates a new user-visible Luna Max thread and sends one self-contained implementation contract with locked decisions, scope, acceptance evidence, and the main-thread handoff target.
4. Sol ends its active turn and stays offline. Luna works in its own thread and contacts Sol only for a real blocker or a terminal handoff.
5. Sol returns on that signal, independently verifies the implementation, handles bounded integration corrections, and accepts the result or requests material rework.

Example implementation brief

Sol does not send Luna a short task summary. It sends one self-contained implementation contract with project grounding, locked decisions, responsibility boundaries, acceptance criteria, communication rules, and the required terminal handoff.

Read a complete implementation brief from an AIWORLDING Character Editor backend task.

Why I use it

I prefer Sol for architecture, ambiguous decomposition, integration, and final acceptance. In my experience, Luna is not reliable enough to own those decisions without strict mediation. I use this protocol to have Sol inspect the project, lock the plan, write complete worker contracts, resolve blockers, and verify every terminal handoff. Luna Max receives only bounded implementation work.

The design goal is to approach the implementation quality I expect from Sol without spending Sol tokens on the entire implementation. Sol spends its budget on architecture, the worker contract, decisions, and verification; Luna Max spends the larger execution budget. This is a workflow goal, not a measured quality guarantee.

The intended token shift looks like this:

Workflow	Illustrative token allocation
Sol-only baseline	approximately 100 million Sol tokens
Subculture target	approximately 50 million Sol tokens and 80+ million Luna tokens

These figures are a planning scenario, not a measured benchmark, fixed ratio, or guarantee. The goal is to reduce Sol execution while retaining Sol control over quality-critical decisions.

That control costs time. Complete contracts, separate task handoffs, code inspection, and stricter verification can make delivery substantially slower than direct single-agent execution.

I also avoid the native Codex subagent lifecycle for this workflow. Subculture requires first-class Codex tasks with their own threadId, visible history, and direct curator handoff. It never substitutes workers that exist only inside a Subagents panel.

Long-running Sol threads

After a long Luna run, a large Sol thread may resume without a useful warm-cache hit and can be expensive to reload. When its context window has grown substantially, I may manually compact the Sol thread before verification. Compaction is optional, and its summary must preserve the locked architecture, ownership map, worker status, acceptance criteria, and unresolved risks.

Requirements

• A Codex app environment that exposes app-level task creation through create_thread.
• Support for creating tasks in the current saved project with the local environment.
• Sol for the main orchestrator and curator role.
• Luna with Max thinking for the default bounded implementation role.

The protocol does not support CLI-only environments that cannot create user-visible tasks. It never falls back to internal spawn_agent, agent-team, or subagent mechanisms.

Install from GitHub

The standard Codex skill installer can install this repository directly. It copies only the subculture skill into your normal Codex skills directory.

PowerShell:

$codexHome = if ($env:CODEX_HOME) { $env:CODEX_HOME } else { Join-Path $HOME ".codex" }
$installer = Join-Path $codexHome "skills/.system/skill-installer/scripts/install-skill-from-github.py"
python $installer --repo FixAdmin/codex-subculture --path skills/subculture

macOS or Linux:

CODEX_HOME="${CODEX_HOME:-$HOME/.codex}"
python3 "$CODEX_HOME/skills/.system/skill-installer/scripts/install-skill-from-github.py" \
  --repo FixAdmin/codex-subculture \
  --path skills/subculture

If a subculture directory already exists, the installer will stop instead of overwriting it. Rename or remove the existing installation before reinstalling.

Install from a local checkout

Copy skills/subculture into your Codex skills directory.

PowerShell:

Copy-Item -Recurse .\skills\subculture "$HOME\.codex\skills\subculture"

macOS or Linux:

cp -R skills/subculture "${CODEX_HOME:-$HOME/.codex}/skills/subculture"

These commands assume that the destination does not already exist. For development, you may link the destination to this checkout so edits take effect without another copy.

Repository layout

skills/subculture/
├── SKILL.md
├── agents/
│   └── openai.yaml
└── references/
    └── implementation-brief.md

SKILL.md contains the orchestration protocol. The bundled reference contains the complete implementation-contract authoring rules and terminal report format. It is installed and versioned with this skill, not as a separate skill.

Validate

python scripts/validate_skill.py

The validator checks the skill metadata, bundled links, mandatory protocol markers, and absence of the retired standalone brief dependency.

Compatibility notes

• Worker creation must return a distinct threadId and produce a separate task in the Codex task list. An entry that appears only under a Subagents panel is not compliant.
• Luna Max is the default implementation worker. Terra and other profiles remain optional, but architecture and final acceptance stay with Sol.
• The exact activation word is SUBCULTURE.

License

This project is released under the MIT License. The same license is included in the installed skill package at skills/subculture/LICENSE.txt.