Claude Code Adapter¶
Adapter ID: claude-code
Source: packages/cli/internal/adapter/claudecode.go
Status: stable
Config directory: .claude
Overview¶
The Claude Code adapter generates native configuration for Claude Code (the claude CLI). It emits agents, skills, commands, output styles, hooks, MCP configuration, and managed settings into .claude/ and .mcp.json.
Generated Files¶
| Path | Description |
|---|---|
.claude/agents/<name>.md |
Agent definitions (canonical agents with Claude Code frontmatter rewrite) |
.claude/skills/<name>/SKILL.md |
Skill directories |
.claude/commands/<name>.md |
Slash commands |
.claude/output-styles/<name>.md |
Output style definitions |
.claude/templates/<name>.md |
Speckit templates |
.claude/hooks/<name>.sh |
Hook scripts (block-destructive-shell, objective-workflow-gate) |
.claude/settings.json |
Managed settings with permissions, hooks config, MCP servers |
.claude/rules/typescript.md |
Sample rule |
.mcp.json |
Project MCP server configuration |
CLAUDE.md |
Root instructions |
Supported Asset Types¶
| Asset kind | Shape | Destination |
|---|---|---|
| Agents | flat | .claude/agents/<name>.md |
| Skills | dir-per-item | .claude/skills/<name>/SKILL.md |
| Templates | flat | .claude/templates/<name>.md |
| Commands | flat | .claude/commands/<name>.md |
| Output styles | flat | .claude/output-styles/<name>.md |
| Chat modes | none | — |
| Prompts | none | — |
MCP Behavior¶
Claude Code MCP is compiled via CompileMCPForTool. At project scope, the adapter writes .mcp.json. At global scope, MCP servers are merged into settings.json (the mcpServers key). When DriveCLI=true, the adapter attempts CLI-driven registration via claude mcp add-json with silent fallback to direct-write.
Hook Behavior¶
Hook scripts are copied from claudecode/hooks/ in the library to .claude/hooks/. The adapter configures them in settings.json under hooks.PreToolUse and hooks.Stop. Two hooks are installed by default:
- block-destructive-shell.sh — PreToolUse matcher for Bash
- objective-workflow-gate.sh — Stop hook
Skill Behavior¶
Skills are written as Agent Skills-compatible directories: .claude/skills/<name>/SKILL.md. The adapter copies selected skills from the canonical library.
Agent Behavior¶
Canonical agents are written as flat markdown files under .claude/agents/. The adapter applies a Claude Code-specific frontmatter rewrite (RewriteAgentForClaudeCode) that strips LazyAI-specific metadata and emits only name and description. A default "guide" agent is always installed. At global scope, legacy flat-layout agents are migrated into the agents/ subdirectory.
Scope Support¶
| Scope | Supported |
|---|---|
| Project | yes |
| Workspace | yes |
| Global | yes |
Headless Support¶
Yes (CanRunHeadless() = true). The adapter supports headless init and validation, including a post-install summary of registered tools.
Known Limitations¶
- No chat modes concept
- No prompts directory; prompts ship as commands
- Global scope MCP is applied via
claude mcp add-json -s user→~/.claude.json(DriveCLI); ifclaudeis not on PATH, no MCP is registered at global scope
Test Coverage¶
| Test file | What it verifies |
|---|---|
claudecode_frontmatter_test.go |
Agent frontmatter rewrite |
claudecode_global_layout_test.go |
Global scope layout |
claudecode_drivecli_test.go |
CLI-driven MCP registration |
claude_cli_test.go |
CLI interaction |