Adapters
An adapter teaches aidokit how to talk to one AI coding CLI. Three first-party adapters ship at v1.0.
Comparison
| Adapter | Conformance | Engine dir | Rules file | Watchdog enforcement |
|---|---|---|---|---|
| claude-code recommended | Strict | .claude/ |
CLAUDE.md |
Hook-level (deterministic) |
| codex | Minimum | .codex/ |
AGENTS.md |
Sandbox + approval policy; prose for watchdog |
| copilot | Minimum | .copilot/ |
.github/copilot-instructions.md |
Prose only (no hook mechanism upstream) |
claude-code (Strict)
The reference adapter. Targets Claude Code ≥ 2.1.32. Emits all seven roles, five output styles, eight schemas, and 19 hook scripts.
Capability profile mapping
| Profile | Mechanism |
|---|---|
read-only | permissionMode: plan |
docs-only-write | permissionMode: default + validate-docs-only.mjs PreToolUse hook |
scope-limited-write | permissionMode: default + validate-scope.mjs PreToolUse hook |
Capability declarations
- writesPaths:
CLAUDE.md,.claude/**,.aido/**,agent-artifacts/**,.gitignore. - runsShellCommands:
claude mcp add,claude mcp remove,claude mcp list,bd init. - networkCalls: none.
Install
npx aidokit@alpha init --adapter claude-code
codex (Minimum)
Targets Codex CLI's stable release surface. Emits the verbs, six roles, and skills, but without hook-level enforcement — watchdog stop conditions are prose-only and rely on model cooperation.
Capability profile mapping
| Profile | Mechanism |
|---|---|
read-only | sandbox: read-only + approval_policy: untrusted |
docs-only-write | sandbox: workspace-write + role-prompt scope discipline |
scope-limited-write | sandbox: workspace-write + per-edit approval |
Documented gaps
- No hook mechanism → watchdog is prose-level.
- No output-style mechanism → Standard's output-style check unreachable.
manifest.gaps. Marketing prompt-level enforcement as deterministic would be a conformance violation.Install
npx aidokit@alpha init --adapter codex
copilot (Minimum)
Targets GitHub Copilot CLI. Emits .github/copilot-instructions.md (auto-loaded workspace-wide by Copilot CLI) plus a .copilot/ tree with agent, prompt, and skill files referenced from the instructions file as documentation.
Documented gaps
- No hook mechanism upstream → watchdog is prose-only.
- No output-style mechanism → Strict's AD-STR-02 check unreachable.
- No native skill auto-discovery → skill files referenced as documentation.
- Partial MCP support →
installMCP/removeMCP/listInstalledMCPsstubbed pending upstream stabilisation.
Install
npx aidokit@alpha init --adapter copilot
Choosing an adapter
| You want… | Use… |
|---|---|
| Deterministic watchdog enforcement | claude-code |
| The full Strict surface (seven roles, output styles) | claude-code |
| To use Codex CLI | codex (Minimum) |
Workspace-wide Copilot integration via copilot-instructions.md | copilot (Minimum) |
| Two or three at once | See Multi-adapter projects |
Adding an adapter to an existing project (v1.0)
npx aidokit add adapter codex --conformance minimum
Shared docs/ and agent-artifacts/ are preserved; only the new adapter's engine directory and rules file are added.
Removing an adapter (v1.0)
npx aidokit remove adapter codex # keeps shared docs by default npx aidokit remove adapter codex --keep-shared=false
Third-party adapters
Post-v1.0, third-party adapters can be published under any npm scope (e.g. @acme/aidokit-adapter-mytool) or unscoped (aidokit-adapter-mytool). The @aidokit/* scope is reserved for first-party. Discovery via:
npx aidokit search adapter
Third-party adapters surface a "Verified" badge if they pass the conformance CI and publish with npm provenance.