claude-ultra

Claude Ultra Plugin

Ultra-capable Claude Code workflow with safety hooks, specialist agents, and productivity commands.

v1.1.0 - All hooks now use Claude Code spec-compliant output formats (Dec 2025).

Quick Install

# Add the marketplace
/plugin marketplace add nategarelik/claude-ultra-plugin

# Install the plugin
/plugin install claude-ultra@claude-ultra-marketplace

# Run setup to verify
/claude-ultra:setup

Alternative: Local Install

git clone https://github.com/nategarelik/claude-ultra-plugin
claude --plugin-dir ./claude-ultra-plugin

Features

Slash Commands

Command	Description
/claude-ultra:setup	Initialize and verify configuration
/claude-ultra:mode	Switch autonomy modes
/claude-ultra:verify	Force web verification of claims
/claude-ultra:plan	Enter planning mode for complex tasks
/claude-ultra:fix	Systematic bug fix workflow
/claude-ultra:parallel	Multi-agent parallel execution
/claude-ultra:review	Code review workflow
/claude-ultra:status	Show session status and token usage
/claude-ultra:compact	Compress context when running low
/claude-ultra:rollback	Git rollback helper

Specialist Agents

Agent	Purpose
@architect	System design and architecture decisions
@code-reviewer	Thorough code review
@debugger	Root cause analysis
@task-worker	Parallel file operations (fast)
@explore	Codebase search and discovery
@security-auditor	Security vulnerability checks
@docs-generator	Documentation generation
@performance-optimizer	Performance analysis
@dependency-auditor	Dependency analysis

Automated Hooks (Spec-Compliant)

All hooks use Claude Code spec-compliant output formats:

Event	Hook	Purpose	Output Format
SessionStart	session-initializer	Load session context	hookSpecificOutput.additionalContext
PreToolUse (Bash)	safety-gate	Block dangerous commands	permissionDecision: "deny"
PreToolUse (Bash)	git-safety-net	Prevent destructive git ops	permissionDecision: "deny"/"ask"
PreToolUse (*)	token-budget-guardian	Monitor context usage	permissionDecision
PostToolUse (Edit/Write)	post-edit-format	Auto-format edited files	additionalContext
PreCompact	context-compression-guard	Archive before compression	stdout
SubagentStop	quality-gate	Validate subagent output	decision: "block" + reason
Stop	session-archiver	Archive session to memory bank	stdout

Hook Output Format Reference

Per official Claude Code docs:

# PreToolUse - Use permissionDecision
{
    "hookSpecificOutput": {
        "hookEventName": "PreToolUse",
        "permissionDecision": "deny",  # or "allow" or "ask"
        "permissionDecisionReason": "Reason shown to Claude"
    }
}

# PostToolUse - Use additionalContext
{
    "hookSpecificOutput": {
        "hookEventName": "PostToolUse",
        "additionalContext": "Info for Claude"
    }
}

# SessionStart - additionalContext only
{
    "hookSpecificOutput": {
        "hookEventName": "SessionStart",
        "additionalContext": "Context injected at start"
    }
}

# Stop/SubagentStop - decision/reason at TOP LEVEL
{
    "decision": "block",
    "reason": "Reason for continuing"
}

Workflow Protocol

Phase 1: UNDERSTAND

• Parse the request
• Check constraints and existing patterns
• Run git status

Phase 2: PLAN

• Use /claude-ultra:plan for complex tasks
• Use @explore before modifying unfamiliar code
• Break work into atomic tasks

Phase 3: EXECUTE

• Delegate to specialist agents
• Use @task-worker for parallel operations
• Keep commits atomic

Phase 4: VERIFY

• Run tests/build/typecheck
• Show output as evidence
• Never claim "done" without proof

Autonomy Modes

/claude-ultra:mode autonomous   # Full autonomy
/claude-ultra:mode collaborative # Check in at decision points (default)
/claude-ultra:mode supervised   # Step-by-step approval

Configuration Examples

• settings.example.json - Optimized settings with all hooks
• CLAUDE.example.md - Slim 66-line system instructions

Compatibility

Works with: superpowers, context7, greptile, LSP plugins

Requirements

• Python 3.8+ (for hooks)
• Claude Code CLI v2.0.12+

License

MIT