Claude Agent Kit

A multi-agent orchestration plugin for Claude Code. Provides persona-based workflows, planning and execution pipelines, specialist subagents, and continuation control using hooks, skills, subagents, and repo-local state files.

It's essentially a port of oh-my-opencode for Claude Code.

Features

• Ultrawork trigger via ulw/ultrawork keyword and /claude-agent-kit:ulw skill.
• Plan to execution flow via /claude-agent-kit:plan and /claude-agent-kit:start-work using .agent-kit/boulder.json.
• Bounded continuation enforcement in Stop hook with max blocks, cooldown, and escape hatch.
• Session resume injection from active boulder state on SessionStart.
• Leaf specialists: explore, librarian, oracle, metis, momus.
• Main-session persona controls: /claude-agent-kit:sisyphus, /claude-agent-kit:hephaestus, /claude-agent-kit:prometheus, /claude-agent-kit:atlas.

Limitations

• Hidden model override is not supported.
• Native custom tools are not supported; use built-in tools and MCP wrappers.
• Nested subagent orchestration is not supported.
• Full tool output rewriting is not supported.

Security and Permissions

Recommended posture (documentation only):

• allow: Read, Grep, Glob
• ask: Bash, Edit, Write, WebFetch, MCP
• deny destructive bash patterns: rm -rf, mkfs, dd if=

The plugin does not ship a settings.json and does not auto-change user permissions.

Compatibility

• Target runtime: Claude Code plugin hooks with SessionStart, UserPromptSubmit, PreToolUse, Stop.
• Dependencies: bash required, jq optional (preferred for strict JSON handling).
• Hook stdin fields can vary by Claude Code version; scripts use fallback extraction and fail-open behavior.

Selftest Pass Criteria

Selftest passes when:

• evidence logs are written under .agent-kit/evidence/<area>/... for each scenario
• plugin validation commands report no errors in a real plugin runtime
• destructive Bash guard scenario is blocked

Smoke Run

1. Enable plugin in Claude Code.
2. Run /plugin validate and /plugin errors.
3. Run /claude-agent-kit:selftest and collect evidence under .agent-kit/evidence/final/.
4. Confirm escape hatch works with /claude-agent-kit:stop-continuation.