Tool Adapters

Codex

Codex reads project skills from .agents/skills/. Invoke them with names such as $discovery, $overview, $feature, $tests, $tests browser, $ci, and $implement, or ask naturally to run the corresponding workflow.

Claude Code

Claude Code reads project skills from .claude/skills/. Invoke them with slash commands such as /discovery, /overview, /feature, /tests, /tests browser, /ci, and /implement.

If Blueprint was installed while Claude Code was already open, restart Claude Code in the project folder so newly added skills are discovered.

New installations import only AGENTS.md on every Claude turn. The project overview, active spec, coding standards, and interaction rules stay in the project and load when the work needs them. Existing projects can run /doctor to identify obsolete direct imports before editing CLAUDE.md.

GitHub Copilot

GitHub Copilot reads AGENTS.md and shares the project skills under .agents/skills/ with Codex. Ask Copilot to run the matching skill, such as the onboard, feature, implement, or check skill.

Google Antigravity

Google Antigravity reads AGENTS.md and shares .agents/skills/ with Codex and GitHub Copilot. Invoke Blueprint skills with slash commands such as /onboard, /feature, or /implement. The installer records Antigravity as a distinct adapter without creating an Antigravity-specific skill tree.

OpenCode

OpenCode reads AGENTS.md and can discover compatible skills under either .agents/skills/ or .claude/skills/. Ask OpenCode to run the matching skill. The installer reuses one of those compatible trees and does not create a duplicate .opencode/skills/ copy.

Select one or more adapters

The interactive installer shows all five tools in a checkbox list, with Codex and Claude Code selected by default. Keep those choices, select another combination, or combine flags for a non-interactive setup:

npx create-ai-blueprint@latest -- --codex --opencode

Available flags are --codex, --claude, --copilot, --antigravity, and --opencode. --all selects every supported tool.

Adapters can be added or removed after installation with npx create-ai-blueprint@latest update. The interactive update shows the same checkbox pre-filled with the installed adapters, and an adapter flag adds a tool without a prompt:

npx create-ai-blueprint@latest update -- --codex

See Change adapters for the removal rules and the shared-tree behavior.

Shared state

All five adapters read and write the same files under blueprint/. Switching tools does not require exporting conversation memory because the current plan, feature checklist, standards, and history live in the repository.

Commit and PR attribution

Blueprint instructs every agent to omit AI co-author trailers, generated-by signatures, and other AI attribution from commits and pull requests while preserving genuine human attribution.

Claude Code also supports optional attribution settings. Merge these keys into the project’s .claude/settings.json or your personal ~/.claude/settings.json, preserving other settings:

{
  "attribution": {
    "commit": "",
    "pr": "",
    "sessionUrl": false
  }
}

Empty strings suppress Claude Code’s default commit trailer and PR attribution text. sessionUrl: false also omits the session link from cloud or Remote Control commits and PRs. The older includeCoAuthoredBy setting is deprecated. Codex’s former commit_attribution TOML key was removed.

The installer does not change tool settings. Instructions and optional settings do not enforce message contents; review the final commit or PR text. Existing installations can manually merge this guidance into their preserved AGENTS.md and blueprint/context/ai-interaction.md files.

Select an independent reviewer

Independent review uses the current runtime automatically by default. When that runtime can start an isolated reviewer with no builder transcript and expose the exact reviewer identity, model, and completion, Blueprint runs the project’s installed Audit skill there. Adapter files alone do not prove that runtime capability.

When automatic execution is unavailable, or review.independentExecution is set to manual, /audit independent current offers only adapters recorded by the Blueprint installation. Selecting an adapter means the project contains that adapter’s files. It does not install, authenticate, or license the external tool itself.

The builder asks which available model should review and recommends an equal-or-stronger coding model, a different model family when practical, and high reasoning for sensitive work. When a tool cannot select a model until the fresh session opens, the receipt records the exact model exposed by that session.

Codex, GitHub Copilot, and Google Antigravity share .agents/. OpenCode can reuse that tree or .claude/, but all four remain separate selections in the installer manifest. Onboard uses that manifest instead of treating a shared tree as every compatible tool. Legacy installations without valid adapter metadata must confirm which tools they use. Claude Code uses .claude/.

Automatic review never searches for global agent roles, prompts, skills, or another workflow. It uses the project-local Audit contract so every installed adapter keeps the same review boundary.

Keep adapters aligned

When changing a shared skill, apply the equivalent behavior to both adapter folders. Codex, GitHub Copilot, and Google Antigravity share .agents/skills/. Claude Code uses .claude/skills/, while OpenCode reuses either compatible tree. Tool-specific invocation wording can differ, but project safety and workflow boundaries should remain equivalent.