cc-settings

AI coding configuration for the Darkroom team. One installer configures Claude Code, Codex, or both with shared standards and workflows plus the native features each product supports.

cc-settings turns either TUI into the team's house engineer. It knows our standards, avoids dangerous actions, proves work before shipping, and delegates larger tasks instead of grinding through them in one context.

macOS / Linux:

bash <(curl -fsSL https://raw.githubusercontent.com/darkroomengineering/cc-settings/main/setup.sh)

Windows (PowerShell):

powershell -ExecutionPolicy Bypass -c "irm https://raw.githubusercontent.com/darkroomengineering/cc-settings/main/setup.ps1 | iex"

The default auto target installs both products when codex is on PATH, and Claude Code only otherwise. Clone the repository to select a target explicitly:

bash setup.sh --target=auto
bash setup.sh --target=claude
bash setup.sh --target=codex
bash setup.sh --target=both

The remote one-liners cannot forward flags. Restart every selected TUI after install. In Codex, review the installed plugin hooks once through /hooks. Flags, requirements, profiles, installed paths, and rollback details are in docs/install.md. Codex-specific boundaries are in docs/codex.md.

There are no commands to learn. Open Claude Code or Codex in a project and describe what you want in your own words. The right workflow loads on its own.

> fix the login redirect
> add a dashboard with stats
> review my changes
> ship it

That's the whole interface. Slash names like /fix and /ship exist so you can pin a specific one when you want to, and you'll rarely need them. When you're curious what fired and why, run bun run whats-on or ask "what skill handles X?". The situation-to-phrase table in MANUAL.md covers the common cases.

• AGENTS.md contains the portable coding standards. Claude Code also receives Claude-specific behavior from CLAUDE.md. Codex receives a marked managed block in its native AGENTS.md, preserving unrelated user text.
• Path-conditioned rules and stack profiles give Claude Code deeper file and framework context. Codex receives its native command policy plus the shared instructions and skills.
• The team knowledge system provides shared conventions, project memory, and /share-learning.

• 38 shared skills fire from ordinary requests such as "fix the login redirect" or "ship it". MANUAL.md maps situations to phrasings.
• Native role agents support planning, implementation, testing, review, security, and orchestration. Claude Code also gets codex-verifier, which is omitted from Codex because that role bridges from Claude Code into Codex.
• Lifecycle hooks enforce proof, safety, and workflow checks. Codex installs the compatible subset through darkroom@cc-settings; review those hooks through /hooks after installation or change.

• Claude Code model pins and routing keep planning, fan-out, and cross-model review on their intended model pools (docs/agent-models.md).
• The Codex bridge lets Claude Code call Codex as a second model family (docs/codex-bridge.md). Standalone Codex workflows never invoke that bridge recursively.
• Claude's custom statusline shows model, context, quota, drift, and bridge health. Codex has no arbitrary custom statusline API, so use Codex /status for session state and installer --status for install health.

Claude Code receives composed settings, permissions, hooks, and four MCP servers: context7, tldr, figma, and chrome-devtools. Codex automatically receives only the fixed HTTPS Figma MCP. It does not auto-run Context7 or Chrome DevTools from unpinned packages, and it does not bundle tldr. Users may configure reviewed and pinned MCP versions themselves.

Claude setup fingerprints its hooks and installed scripts, then verifies them on session start. bun run audit:hooks classifies hook commands against trusted content and suspicious patterns (SECURITY.md). Codex uses its native plugin trust review through /hooks.

The installer supports --target=auto|claude|codex|both for install, dry run, status, rollback, and uninstall. Backups and ownership records are product-specific. Reinstalls preserve unrelated Codex agents, rules, and AGENTS.md text, plus unrelated Claude settings and extensions. Invalid or unknown flags fail closed. Codex copies an allowlisted runtime package, not user or ignored checkout files.

Everything is TypeScript on Bun, tested, linted, and CI-gated, including the shared counts and installer contracts above.

Rules start as advisories, work ships through gates, and an advisory is only promoted to a gate when telemetry proves it gets ignored. The full loop, with diagram: docs/the-flow.md.

Shared config means every engineer on the team runs the same agents, the same guardrails, and the same coding standards — no per-machine drift. AGENTS.md is the open standard for AI coding instructions and is also read by Codex, Cursor, Copilot, and Windsurf, so the rules follow you across tools.

darkroom.engineering · MIT