Claude Runtime

April 30, 2026 · View on GitHub

The Claude Code plugin lives in the claude/ subdirectory.

Configuration

Manifest: claude/.claude-plugin/plugin.json Version: generated from package.json Hooks: claude/hooks/claude-hooks.json MCP Config: claude/.mcp.json

MCP Server

{
  "mcpServers": {
    "maestro": {
      "command": "node",
      "args": ["${CLAUDE_PLUGIN_ROOT}/mcp/maestro-server.js"],
      "cwd": "${CLAUDE_PLUGIN_ROOT}"
    }
  }
}

The public server at claude/mcp/maestro-server.js is a thin adapter. It sets MAESTRO_RUNTIME=claude, prefers canonical repo src/mcp/maestro-server.js when the plugin is loaded from a checkout, and falls back to bundled claude/src/mcp/maestro-server.js for detached marketplace or copied installs. Claude declares primary: filesystem and fallback: none.

Agent Naming

Claude uses kebab-case for agent names: code-reviewer, api-designer, accessibility-specialist.

Agent files are generated at claude/agents/*.md with kebab-case filenames.

Delegation

Claude Code Agent subagent syntax:

Agent(subagent_type: "maestro:coder", prompt: "...")
Agent(subagent_type: "maestro:architect", prompt: "...")

Skills

19 Markdown skills in claude/skills/:

Core (3) — generated public entry points from the core command registry:

orchestrate/SKILL.md
execute/SKILL.md
resume-session/SKILL.md

Entry-point (9) — from registry (review, debug, resume are remapped to review-code, debug-workflow, resume-session because they collide with Claude Code built-in commands; see src/generator/entry-point-expander.js HOST_RESERVED_NAMES):

review-code/SKILL.md, debug-workflow/SKILL.md, archive/SKILL.md, status/SKILL.md
security-audit/SKILL.md, perf-check/SKILL.md, seo-audit/SKILL.md
a11y-audit/SKILL.md, compliance-check/SKILL.md

Infrastructure (7) — discovery stubs for shared skills:

code-review/SKILL.md, delegation/SKILL.md, design-dialogue/SKILL.md
execution/SKILL.md, implementation-planning/SKILL.md
session-management/SKILL.md, validation/SKILL.md

Infrastructure skills have user-invocable: false — they surface in Claude's skill picker but delegate to MCP for content.

Hooks

4 hook entries across 3 event types (2 with matchers):

Event	Matcher	Script	Timeout
`SessionStart`	—	`scripts/hook-runner.js claude session-start`	10s
`PreToolUse`	`Agent`	`scripts/hook-runner.js claude before-agent`	10s
`PreToolUse`	`Bash`	scripts/policy-enforcer.js	5s
`SessionEnd`	—	`scripts/hook-runner.js claude session-end`	10s

Key Differences from Gemini

Uses PreToolUse with matchers instead of BeforeAgent/AfterAgent
No AfterAgent hook — Claude does not validate handoff report format via hooks
Same effective timeouts (10s) but Claude config uses seconds, Gemini uses milliseconds (10000ms)
Policy enforcement via hook script, not TOML

Hook Adapter

claude/scripts/adapters/claude-adapter.js normalizes Claude Code stdin:

Claude Field	Internal Field
`session_id`	`sessionId`
`cwd`	`cwd`
`hook_event_name`	`event`
`tool_input.subagent_type`	`agentName`
`tool_input.prompt`	`agentInput`
`tool_result`	`agentResult`
(hardcoded false)	`stopHookActive`

Output format: { continue: boolean, decision: "approve"|"block", systemMessage?: string, reason?: string }

Policy Enforcement

claude/scripts/policy-enforcer.js — JavaScript-based, triggered on Bash tool use:

Deny rules (blocks command):

rm -rf, rm -fr, sudo rm -rf, sudo rm -fr
git reset --hard, git checkout --
git clean -fd, git clean -df, git clean -xfd, git clean -xdf
Heredocs (<<)

Ask rules (requires user confirmation):

tee commands
Output redirection (>, >>)

Parses compound commands (;, &&, ||, |) and recursively checks subshells ($(...), backticks).

Tool Mapping

Canonical	Claude
`read_file` / `read_many_files`	`Read`
`write_file`	`Write`
`replace`	`Edit`
`list_directory` / `glob`	`Glob`
`grep_search`	`Grep`
`google_web_search`	`WebSearch`
`web_fetch`	`WebFetch`
`run_shell_command`	`Bash`
`ask_user`	`AskUserQuestion`
`write_todos`	`TaskCreate`, `TaskUpdate`, `TaskList`
`activate_skill`	`Skill`
`enter_plan_mode`	`EnterPlanMode`
`exit_plan_mode`	`ExitPlanMode`
`codebase_investigator`	`Agent (Explore)` / `Grep` / `Glob`

Feature Flags

The canonical feature set (same 4 flags across all runtimes, values per runtime):

exampleBlocks:             true   (examples embedded in description)
claudeStateContract:       true   (Claude-specific session-state contract)
scriptBasedStateContract:  false
codexStateContract:        false

See src/platforms/claude/runtime-config.js for the authoritative values.

Agent Frontmatter

Claude agent stubs include example blocks in the description and use camelCase for turns:

---
name: coder
description: |
  Implementation specialist...
  
  <example>
  Context: User needs a new feature implemented.
  user: "Implement the login flow"
  assistant: "I'll follow existing patterns..."
  </example>
model: inherit
color: green
maxTurns: 25
tools:
  - Read
  - Write
  - Edit
  - Bash
  - Glob
  - Grep
---

Fields: model (always "inherit"), color, maxTurns (camelCase). No temperature or timeout.

Generated Files

claude/
├── agents/                39 agent stubs (kebab-case)
├── skills/                19 skill directories
├── hooks/                 1 hook config (claude-hooks.json)
├── scripts/               thin hook wrapper, adapter wrapper, policy enforcer
├── mcp/                   thin MCP entrypoint
├── src/                   generated detached runtime payload for isolated installs
├── .claude-plugin/        1 plugin manifest
├── .mcp.json
├── mcp-config.example.json
└── README.md