Globales Ranking · von 601 Skills
agent-security-engineer Hermes AI Agent Skill
Quellcode ansehen: tony363/superclaude
MediumInstallation
npx skills add tony363/superclaude --skill agent-security-engineer 77
Installationen
SuperClaude Framework
A config-first meta-framework for Claude Code that provides 19 specialized agent personas (5 core + 7 traits + 7 extensions), 27 structured commands, 43 skills, and comprehensive MCP integration with quality-driven iterative workflows.
SuperClaude transforms Claude Code into a powerful development platform with specialized agent prompts, signal-based loop orchestration, and multi-model consensus capabilities. The core interface is markdown and YAML configuration files, with a Python orchestration layer for advanced workflows including quality gates, termination detection, and PAL MCP integration.
Table of Contents
- Overview
- Key Features
- Architecture
- Core Orchestration Layer
- Core API Reference
- Installation
- Quick Start
- Dashboard
- Agent System
- Command System
- Loop Orchestration
- MCP Integrations
- Skills System
- Quality System
- Quality Gates
- Metrics System
- Type System
- Configuration
- Directory Structure
- Creating Custom Agents
- Contributing
Overview
SuperClaude is a meta-prompt framework that enhances Claude Code with:
- 19 Specialized Agents: 5 core + 7 composable traits + 7 domain extensions (tiered architecture)
- 43 Active Skills: 8 agent personas + 26 command workflows + 6 GitNexus code intelligence + 3 utility skills
- 27 Structured Commands: analyze, implement, test, design, document, and more
- 6 Framework Modes: normal, brainstorming, introspection, task_management, token_efficiency, orchestration
- MCP Integration: PAL (11 tools), Rube (500+ apps via Composio, including web search)
- Quality Gates: KISS, Purity, SOLID, and Let It Crash validators with iterative quality loop
- Core Orchestration: ~2,800 lines Python for loop management, PAL integration, and skill learning
- Signal-Based Architecture: Structured communication between components
- Metrics System: Callback-based operational metrics with Prometheus/StatsD integration
Key Features
Config-First Hybrid Architecture
SuperClaude v7.0.0 is a config-first hybrid framework:
- Markdown Agent Personas: Each agent is a self-contained markdown file with YAML frontmatter
- YAML Configuration: 6 configuration files for agents, commands, quality, models, MCP, and framework settings
- Python Orchestration: Loop orchestrator, quality assessment, PAL integration, and skill learning
- Portable: Works with any Claude Code instance
- Extensible: Add agents by creating markdown files
Why Config-First Hybrid?
| Benefit | Description |
|---|---|
| Simplicity | Core interface is markdown/YAML files |
| Portability | Works with any Claude Code instance |
| Extensibility | Add agents by creating markdown files |
| Maintainability | Prompts are easy to refine and version |
| Version Control | Easy to diff and review prompt changes |
| Advanced Workflows | Python orchestration for quality gates and loop control |
| Safety | Hard limits on iterations, termination detection |
Tiered Agent Architecture
19 agents organized in a tiered system for composable expertise:
| Tier | Count | Purpose | Examples |
|---|---|---|---|
| Core | 5 | High-priority generalists | architect, developer, optimizer, guardian, communicator |
| Traits | 7 | Composable modifiers | security-first, performance-first, test-driven, minimal-changes, principles-enforced, mcp-pal-enabled, mcp-rube-enabled |
| Extensions | 7 | Domain specialists | typescript-react-expert, python-expert, golang-expert, rust-expert, kubernetes-specialist, data-engineer, ml-engineer |
Why Tiered? The v7 architecture provides a lean, composable system. Core agents handle most tasks, traits modify behavior (e.g., @architect +security-first), and extensions provide deep domain expertise when needed.
Command System
27 structured commands with consistent patterns:
/sc:analyze - Static analysis, security review, performance bottlenecks
/sc:implement - Feature implementation with quality gates and loop support
/sc:test - Test execution, coverage analysis, and test generation
/sc:design - Architecture and system design with ADRs
/sc:document - Documentation generation and maintenance
/sc:brainstorm - Creative ideation and exploration
/sc:explain - Educational explanations and learning
/sc:improve - Code improvement, refactoring, optimization
/sc:build - Build system and compilation workflows
/sc:git - Git operations, smart commits, branch management
/sc:workflow - Multi-step task orchestration
/sc:estimate - Effort estimation and planning
/sc:cicd-setup - CI/CD workflow and pre-commit generation
/sc:readme - Auto-update README.md from git diff with PAL consensus
/sc:pr-fix - Create PR and iteratively fix CI failures
/sc:pr-check - Pre-PR local CI validation with auto-fix
/sc:log-fix - Log analysis and iterative bug fixing
/sc:e2e - E2E testing (Playwright, Cypress, Selenium)
/sc:tdd - Test-driven development with Red-Green-Refactor
/sc:principles - KISS, Purity, SOLID, Let It Crash validation
/sc:worktree - Git worktree management for parallel development
/sc:mcp - MCP orchestration hub (PAL + Rube)
/sc:code-review - Multi-model consensus code review
/sc:research - Deep research with web search + consensus
/sc:eda - Exploratory data analysis and visualization
/sc:evaluate - LLM pipeline evaluation with judge scoring
/sc:push - Multi-remote git push with selective filteringArchitecture
How SuperClaude Works
flowchart TB
subgraph User["User Layer"]
REQ["User Request<br/>/sc:implement --loop"]
end
subgraph Runtime["Claude Code Runtime"]
CLAUDE["CLAUDE.md<br/>Master System Prompt"]
SELECT["Agent Selection<br/>Semantic Matching + Weights"]
EXEC["Task Execution"]
end
subgraph Config["Configuration Layer"]
AGENTS[("agents/<br/>5 core + 7 traits + 7 ext")]
COMMANDS[("commands/<br/>16 templates")]
SKILLS[(".claude/skills/<br/>38 skills")]
YAML[("config/<br/>6 YAML files")]
end
subgraph Core["Core Orchestration (Python)"]
direction TB
LOOP["LoopOrchestrator<br/>Max 5 iterations"]
QUALITY["QualityAssessor<br/>9 dimensions"]
TERM["Termination Detection<br/>5 conditions"]
PALINT["PAL Integration<br/>Signal generation"]
LEARN["Skill Learning<br/>Pattern extraction"]
end
subgraph MCP["MCP Integrations"]
PAL["PAL MCP<br/>11 tools"]
RUBE["Rube MCP<br/>500+ apps"]
end
REQ --> CLAUDE
CLAUDE --> SELECT
SELECT --> AGENTS
SELECT --> COMMANDS
SELECT --> EXEC
EXEC --> SKILLS
EXEC <--> Core
LOOP --> QUALITY
QUALITY --> TERM
TERM --> PALINT
PALINT <--> PAL
EXEC <--> RUBE
Config --> Runtime
style User fill:#e8f5e9
style Runtime fill:#e3f2fd
style Config fill:#fff3e0
style Core fill:#fce4ec
style MCP fill:#f3e5f5File-Based Configuration
SuperClaude uses a layered file-based architecture:
- CLAUDE.md - Master system prompt loaded by Claude Code
- agents/index.yaml - Agent registry with triggers, categories, and selection weights
- agents/core/*.md - 5 core agent persona prompts
- agents/traits/*.md - 7 composable behavior modifier prompts
- agents/extensions/*.md - 7 domain specialist prompts
- commands/index.yaml - Command registry with flags and aliases
- commands/*.md - 16 command templates
- config/*.yaml - 6 configuration files
- core/*.py - Python orchestration modules
- mcp/*.md - MCP integration guides
- .claude/skills/ - 43 Claude Code skills
Core Orchestration Layer
SuperClaude v7.0.0 includes a Python orchestration layer in core/ for advanced workflows:
Modules
| Module | Lines | Purpose |
|---|---|---|
loop_orchestrator.py |
~410 | Manages iterative improvement with quality gates |
quality_assessment.py |
~265 | 9-dimension quality scoring with evidence collection |
pal_integration.py |
~175 | PAL MCP signal generation (review, debug, validation) |
types.py |
~150 | Core type definitions (TerminationReason, LoopConfig, etc.) |
metrics.py |
~225 | Callback-based metrics protocol and emitters |
skill_learning_integration.py |
~550 | Skill extraction from successful executions |
skill_persistence.py |
~980 | Skill storage, retrieval, and promotion |
LoopOrchestrator
The LoopOrchestrator class manages the --loop flag behavior:
stateDiagram-v2
[*] --> Initialize
state "Loop Execution" as Loop {
[*] --> ExecuteSkill
ExecuteSkill --> CollectEvidence
CollectEvidence --> AssessQuality
AssessQuality --> CheckThreshold
state CheckThreshold <<choice>>
CheckThreshold --> Success: score >= 70
CheckThreshold --> CheckTermination: score < 70
state CheckTermination <<choice>>
CheckTermination --> MaxIterations: iteration >= max
CheckTermination --> Timeout: wall-clock exceeded
CheckTermination --> Error: skill execution failure
CheckTermination --> Continue: OK to iterate
Continue --> PALReview: Generate signal
PALReview --> PrepareNext
PrepareNext --> ExecuteSkill
}
Initialize --> Loop
Success --> FinalValidation: PAL final signal
MaxIterations --> [*]
Timeout --> [*]
Error --> [*]
FinalValidation --> [*]
note right of Initialize
Triggered by /sc:implement --loop
end noteKey Features:
- Safety: Hard maximum of 5 iterations (cannot be overridden via
hard_max_iterations) - Quality-Driven: Stops when score >= 70 (configurable threshold)
- PAL Integration: Generates review signals within loop for external validation
- Signal-Based: Communicates with Claude Code via structured signals
- Evidence Collection: Tracks changes, tests, lint results, and file modifications
Signal Architecture
The orchestrator uses structured signals for component communication:
# Skill invocation signal
{
"action": "execute_skill",
"skill": "sc-implement",
"parameters": {
"task": "description",
"improvements_needed": [...],
"iteration": 2,
"focus": "remediation" # "implementation" for iter 0
},
"collect": ["changes", "tests", "lint", "changed_files"],
"context": {...}
}Core API Reference
Complete API documentation for the SuperClaude Python orchestration layer (core/ module).
Module Exports
from core import (
# Types
TerminationReason, # Enum: 5 loop termination conditions
LoopConfig, # Dataclass: Loop configuration
LoopResult, # Dataclass: Final loop result
IterationResult, # Dataclass: Single iteration result
QualityAssessment, # Dataclass: Quality score + improvements
# Classes
QualityAssessor, # Quality scoring wrapper
PALReviewSignal, # PAL MCP signal generator
LoopOrchestrator, # Main loop controller
)LoopOrchestrator
The main controller for --loop functionality.
class LoopOrchestrator:
"""
Lightweight agentic loop orchestrator.
Thread Safety: NOT thread-safe. Create new instance per task/thread.
"""
def __init__(
self,
config: Optional[LoopConfig] = None,
logger: Optional[logging.Logger] = None,
metrics_emitter: Optional[MetricsEmitter] = None,
) -> None:
"""
Initialize the loop orchestrator.
Args:
config: Loop configuration (defaults to LoopConfig())
logger: Logger instance for structured logging
metrics_emitter: Callback for operational metrics
"""
def run(
self,
initial_context: dict[str, Any],
skill_invoker: Callable[[dict[str, Any]], dict[str, Any]],
) -> LoopResult:
"""
Execute the agentic loop.
Args:
initial_context: Task context with:
- task: Description of what to implement
- improvements_needed: Initial improvements (optional)
- changed_files: Already modified files (optional)
skill_invoker: Function that invokes Skills via Claude Code
Should return dict with: changes, tests, lint, changed_files
Returns:
LoopResult with final output, assessment, and history
"""Usage Example:
from core import LoopOrchestrator, LoopConfig
config = LoopConfig(
max_iterations=3,
quality_threshold=70.0,
pal_review_enabled=True,
)
orchestrator = LoopOrchestrator(config)
def my_skill_invoker(context):
return {
"changes": [...],
"tests": {"ran": True, "passed": 10, "failed": 0},
"lint": {"ran": True, "errors": 0},
"changed_files": ["src/module.py"],
}
result = orchestrator.run(
initial_context={"task": "Implement user authentication"},
skill_invoker=my_skill_invoker,
)
print(f"Final score: {result.final_assessment.overall_score}")
print(f"Termination: {result.termination_reason.value}")QualityAssessor
Wraps evidence_gate.py for quality scoring.
class QualityAssessor:
def __init__(self, threshold: float = 70.0) -> None
def assess(self, context: dict[str, Any]) -> QualityAssessmentScoring Breakdown (Fallback Mode):
| Component | Points | Condition |
|---|---|---|
| File changes | 30 | Any changes detected |
| Tests executed | 25 | Tests ran |
| Tests passing | 20 | All tests pass |
| Lint clean | 15 | No lint errors |
| Coverage | 10 | 80%+ coverage |
PALReviewSignal
Generates signals for PAL MCP invocation.
class PALReviewSignal:
TOOL_CODEREVIEW = "mcp__pal__codereview"
TOOL_DEBUG = "mcp__pal__debug"
TOOL_THINKDEEP = "mcp__pal__thinkdeep"
TOOL_CONSENSUS = "mcp__pal__consensus"
@staticmethod
def generate_review_signal(
iteration: int,
changed_files: list[str],
quality_assessment: QualityAssessment,
model: str = "gpt-5",
review_type: str = "auto", # auto/quick/full
) -> dict[str, Any]
@staticmethod
def generate_debug_signal(
iteration: int,
termination_reason: str,
score_history: list[float],
model: str = "gpt-5",
) -> dict[str, Any]
@staticmethod
def generate_final_validation_signal(
changed_files: list[str],
quality_assessment: QualityAssessment,
iteration_count: int,
model: str = "gpt-5",
) -> dict[str, Any]Helper Functions
# Create skill invocation signal
def create_skill_invoker_signal(context: dict[str, Any]) -> dict[str, Any]
# Merge PAL feedback into context
def incorporate_pal_feedback(
context: dict[str, Any],
pal_result: dict[str, Any],
) -> dict[str, Any]
# Convenience function for quality assessment
def assess_quality(
context: dict[str, Any],
threshold: float = 70.0
) -> QualityAssessmentInstallation
Option 1: Clone Repository
git clone https://github.com/Tony363/SuperClaude.git
cd SuperClaudeOption 2: Add as Git Submodule
git submodule add https://github.com/Tony363/SuperClaude.git SuperClaudeSetup Claude Code
Add to your project's .claude/settings.json:
{
"systemPromptFiles": ["SuperClaude/CLAUDE.md"]
}Or copy CLAUDE.md to your project root.
Optional: Python Dependencies
For advanced workflows using the core orchestration layer:
pip install -e .Security Layer: Sondera Integration
SuperClaude supports optional integration with Sondera, a security harness that validates Claude Code tool executions using LLM-powered policy enforcement and Cedar policy language.
Important: Sondera is not bundled by default. The standard installation (above) does not include it. Sondera must be explicitly installed when needed for production or compliance requirements.
📖 Full integration guide and architecture →
Default Install (Development)
The standard installation provides full SuperClaude functionality without security restrictions. This is ideal for:
- Local development and experimentation
- Trusted single-user environments
- Research and learning
- Scenarios where performance is critical
Production Setup (With Sondera)
For production deployments or high-security environments, install with Sondera:
./install-with-sondera.shWhat Sondera provides:
- Pre-execution validation of core Claude Code tools (Read, Write, Edit, Bash, etc.)
- LLM-powered intent analysis (ministral-3:14b-cloud)
- Safety validation (gpt-oss-safeguard:20b)
- Cedar policy enforcement engine
- Audit logging of all validated operations
Prerequisites:
- Rust toolchain (
cargo) - Ollama with models:
ministral-3:14b-cloud,gpt-oss-safeguard:20b(~30GB) - Cedar policy engine (embedded in harness)
- Unix socket support (Linux/macOS only; Windows not supported)
Testing the integration:
./test-sondera-integration.shTradeoffs:
| Aspect | Without Sondera | With Sondera |
|---|---|---|
| Installation | Simple (git clone) | Complex (Rust, Ollama, models) |
| Performance | Fast | Slower (LLM validation overhead) |
| Security | Trust-based | Policy-enforced |
| Use Case | Development, research | Production, compliance |
| Maintenance | Low | Higher (harness service) |
| Platform | Cross-platform | Linux/macOS only |
Disabling Sondera:
# Remove hook configuration
rm .claude/settings.local.json
# Stop harness service
# If using systemd (Linux):
systemctl --user stop sondera-harness
# If launched manually:
pkill -f sondera-harnessFailure Mode: When the harness service is unavailable, tool calls fail-closed (blocked by default) for safety. Configure fail-open mode in the harness settings if needed.
Quick Start
1. Natural Language
User: Help me debug this authentication issue
Claude: [Selects communicator agent, applies debugging methodology]2. Using Commands
User: /sc:analyze src/auth/
Claude: [Runs comprehensive static analysis, security review, quality assessment]3. Specifying Agents
User: @python-expert Review this Flask application
Claude: [Uses Python expert persona with framework-specific knowledge]4. Iterative Improvement
User: /sc:implement new user registration --loop
Claude: [Implements with quality gates, iterates until score >= 70 or max 5 iterations]5. Using MCP Tools
With PAL MCP configured:
User: Use PAL to review this code
Claude: [Invokes mcp__pal__codereview for multi-model code review]Dashboard
SuperClaude includes a native desktop dashboard for visualizing and controlling the agentic framework.
Features
- Feature Inventory - Browse all 19 agents, 24 commands, 38 skills, and 6 behavioral modes
- Live Monitor - Real-time execution tracking with event streaming, heartbeat indicator, and quality score visualization
- Execution Control - Start, stop, pause, and resume executions with configurable parameters; expandable detail panel with 5 tabs (Run Instructions, Execution Log, Files Changed, Quality Breakdown, Execution Tree)
- Execution Tree - Visual tree of iterations, tool calls, and subagent spawns built from streaming events
- Diff Viewer - Inline diff display for Edit/Write tool invocations showing before/after changes
- Historical Metrics - View past session data, event timelines, and performance trends
Quick Start
# Terminal 1: Start the daemon
cargo run -p superclaude-daemon
# Terminal 2: Launch the dashboard
cd crates/dashboard
cargo tauri devThe dashboard window will open showing the Feature Inventory page. Navigate using the sidebar to access Monitor, Control, and History views.
Building for Production
cd crates/dashboard
cargo tauri build
# Output packages:
# - Debian/Ubuntu: target/release/bundle/deb/superclaude-dashboard_*.deb
# - Universal Linux: target/release/bundle/appimage/superclaude-dashboard_*.AppImageSystem Requirements
- OS: Linux (tested on Manjaro/Arch, Ubuntu 22.04+)
- Dependencies: webkit2gtk-4.1, libappindicator-gtk3, librsvg
- Runtime: SuperClaude daemon running on port 50051
See Dashboard README for detailed usage guide.
Agent System
Core Agents (5)
High-priority generalists for common tasks:
| Agent | Triggers | Purpose |
|---|---|---|
| architect | architecture, design, system, backend, frontend, api, database, infrastructure, ci/cd, deploy, fullstack | System design spanning architecture, infrastructure, backend, frontend, and deployment |
| developer | implement, create, build, develop, task, general, requirement, spec | Versatile implementation and multi-domain coordination |
| optimizer | optimize, performance, slow, refactor, improve, clean | Performance optimization and code quality improvement |
| guardian | security, vulnerability, test, validate, quality, qa | Security analysis and comprehensive testing |
| communicator | document, docs, explain, learn, tutorial, debug, investigate | Documentation, teaching, and debugging explanations |
Traits (7)
Composable behavior modifiers that can be combined with any agent:
| Trait | Purpose | Example Usage |
|---|---|---|
| security-first | Prioritize security considerations | @architect +security-first |
| performance-first | Focus on optimization | @optimizer +performance-first |
| test-driven | Enforce TDD practices | @developer +test-driven |
| minimal-changes | Reduce change scope | @optimizer +minimal-changes |
| principles-enforced | Enforce SOLID, Let It Crash, KISS, Pure Functions | @architect +principles-enforced |
| mcp-pal-enabled | Enables PAL MCP tools (consensus, debug, codereview) | @guardian +mcp-pal-enabled |
| mcp-rube-enabled | Enables Rube MCP tools (500+ app integrations) | @architect +mcp-rube-enabled |
Extensions (7)
Domain-specialist agents for deep expertise:
| Extension | Triggers | Specialty |
|---|---|---|
| typescript-react-expert | typescript, ts, node, react, hooks, nextjs | TypeScript and React with strong typing |
| python-expert | python, django, fastapi, pytest | Python development with modern patterns |
| golang-expert | go, golang, goroutine | Go development with idiomatic patterns |
| rust-expert | rust, cargo, ownership | Rust with ownership patterns |
| kubernetes-specialist | kubernetes, k8s, helm, pod | Kubernetes orchestration |
| data-engineer | data, pipeline, etl, spark, kafka | Data engineering and pipelines |
| ml-engineer | ml, machine learning, pytorch, tensorflow | Machine learning systems |
Agent Selection Algorithm
Agents are selected using weighted semantic matching:
flowchart TD
START["User Request"] --> PARSE["Parse Keywords & Context"]
PARSE --> TRIGGER{"Exact Trigger<br/>Match?"}
TRIGGER -->|"Yes (0.35)"| CORE{"Core<br/>Agent?"}
TRIGGER -->|"No"| CATEGORY{"Category<br/>Match?"}
CORE -->|"Yes"| USE_CORE["Load Core Agent<br/>(5 available)"]
CORE -->|"No"| USE_EXT["Load Trait/Extension<br/>(14 available)"]
CATEGORY -->|"Yes (0.25)"| CAT_SELECT["Select from Category"]
CATEGORY -->|"No"| DESC{"Description<br/>Match?"}
DESC -->|"Yes (0.20)"| DESC_SELECT["Score by Description"]
DESC -->|"No"| TOOL{"Tool<br/>Match?"}
TOOL -->|"Yes (0.20)"| TOOL_SELECT["Match Required Tools"]
TOOL -->|"No"| FILE{"File<br/>Context?"}
FILE -->|".py files"| PYTHON["Python Expert"]
FILE -->|".ts/.js"| TS["TypeScript Pro"]
FILE -->|".go files"| GO["Golang Pro"]
FILE -->|".rs files"| RUST["Rust Engineer"]
FILE -->|"Other"| GENERAL["General Purpose"]
USE_CORE --> EXEC["Execute Task"]
USE_EXT --> EXEC
CAT_SELECT --> EXEC
DESC_SELECT --> EXEC
TOOL_SELECT --> EXEC
PYTHON --> EXEC
TS --> EXEC
GO --> EXEC
RUST --> EXEC
GENERAL --> EXEC
style START fill:#4caf50,color:#fff
style EXEC fill:#2196f3,color:#fffSelection Weights (from select_agent.py):
selection:
weights:
keyword_match: 0.35 # Keyword/trigger matching (highest priority)
category_match: 0.25 # Domain/category alignment
task_match: 0.20 # Task text matching
file_patterns: 0.10 # File pattern matching
priority_bonus: 0.10 # Tier priority bonus
thresholds:
minimum_score: 0.3 # Minimum match score
confidence_levels:
excellent: 0.7 # >= 0.7 score
high: 0.5 # >= 0.5 score
medium: 0.3 # >= 0.3 scoreCommand System
Commands provide structured execution patterns with optional quality gates.
Command Format
Each command is a markdown file with YAML frontmatter:
---
name: implement
description: Feature implementation with quality gates
aliases: [build, create, develop]
flags: [--loop, --tests, --docs, --pal-review, --consensus]
evidence_required: true
---
# Implement Command
[Structured implementation methodology...]Command Execution Flow
sequenceDiagram
participant U as User
participant C as Claude Code
participant CMD as Command Template
participant A as Agent Persona
participant O as LoopOrchestrator
participant Q as QualityAssessor
participant P as PAL MCP
U->>C: /sc:implement feature --loop
C->>CMD: Load implement.md template
CMD->>A: Activate relevant personas
C->>O: Initialize loop (max 5 iter)
rect rgb(230, 245, 255)
note over O,P: Iterative Improvement Loop
loop Iteration 0..N (max 5)
O->>A: Execute with context
A-->>O: Return evidence
O->>Q: Assess quality
Q-->>O: Score + improvements
alt Score >= 70
O-->>C: QUALITY_MET
else Oscillation/Stagnation
O->>P: Generate debug signal
O-->>C: Terminate
else Continue
O->>P: Generate review signal
P-->>O: PAL feedback
O->>A: Next iteration
end
end
end
O->>P: Final validation signal
P-->>O: Validation result
O-->>C: LoopResult
C->>U: Implementation + quality reportAvailable Commands
| Command | Aliases | Evidence | Key Flags |
|---|---|---|---|
| analyze | analyse, check, inspect | No | --security, --deep, --performance |
| implement | build, create, develop | Yes | --loop, --pal-review, --consensus, --tests |
| test | verify, check | Yes | --coverage, --unit, --integration |
| design | architect, plan | No | --adr, --diagram, --review |
| document | docs, readme | No | --api, --readme, --changelog |
| brainstorm | ideate, explore | No | --divergent, --constraints |
| explain | teach, learn, understand | No | --verbose, --simple, --diagram |
| improve | refactor, enhance, optimize | Yes | --performance, --readability, --security |
| build | compile, package | No | --watch, --production, --docker |
| git | commit, branch | No | --commit, --pr, --branch |
| workflow | pipeline, sequence | No | --spec, --parallel |
| estimate | scope, size | No | --breakdown, --risks |
| cicd-setup | cicd-init, pipeline-setup | No | --lang, --minimal, --full |
| readme | readme-sync, doc-sync | No | --base, --preview, --consensus |
| pr-fix | prfix, cifix, fix-pr | Yes | --branch, --auto-fix, --max-fix-attempts |
| pr-check | preflight, prepr | No | --fix, --strict |
| log-fix | logfix, debug-logs | No | --source, --pattern |
| e2e | e2e-test, browser-test | Yes | --browser, --headed, --trace |
| tdd | test-driven | Yes | --cycle, --framework |
| principles | validate-principles | No | --kiss, --solid, --purity, --crash |
| worktree | git-worktree | No | --name, --branch |
| mcp | mcp-orchestrate | No | --pal, --rube, --consensus |
| code-review | review, cr | No | --models, --consensus |
| research | deep-research | No | --depth, --sources |
| eda | data-analysis | No | --format, --output |
| evaluate | eval, judge | No | --dataset, --judge-model |
| push | git-push | No | --remotes, --exclude |
Loop Orchestration
The loop orchestration system manages iterative improvement workflows with safety guarantees.
Termination Conditions
flowchart TD
START["Start Loop Iteration 0"] --> EXEC["Execute Skill"]
EXEC --> EVIDENCE["Collect Evidence"]
EVIDENCE --> ASSESS["Assess Quality"]
ASSESS --> SCORE{"Score >= 70?"}
SCORE -->|"Yes"| QUALITY_MET["QUALITY_MET"]
SCORE -->|"No"| MAX{"Iteration >= max?"}
MAX -->|"Yes"| MAX_ITER["MAX_ITERATIONS"]
MAX -->|"No"| TMOUT{"Timeout?"}
TMOUT -->|"Yes"| TIMEOUT_TERM["TIMEOUT"]
TMOUT -->|"No"| PAL["Generate PAL Review Signal"]
PAL --> FEEDBACK["Incorporate Feedback"]
FEEDBACK --> NEXT["Prepare Next Context"]
NEXT --> EXEC
QUALITY_MET --> FINAL["Final PAL Validation Signal"]
FINAL --> DONE["Return LoopResult"]
MAX_ITER --> DONE
TIMEOUT_TERM --> DONE
style QUALITY_MET fill:#4caf50,color:#fff
style MAX_ITER fill:#ff9800,color:#fff
style TIMEOUT_TERM fill:#f44336,color:#fffAll 5 Termination Reasons
| Reason | Trigger | PAL Signal |
|---|---|---|
QUALITY_MET |
Score >= threshold (70) | Final validation |
MAX_ITERATIONS |
Iteration >= 5 (hard cap) | None |
ERROR |
Skill execution failure | None |
HUMAN_ESCALATION |
Requires human review | None |
TIMEOUT |
Wall-clock time exceeded | None |
PAL Signal Types
The loop orchestrator generates two types of PAL signals:
- Review Signal - Within-loop review for quality improvement
- Final Validation Signal - After successful completion (score >= 70)
MCP Integrations
SuperClaude integrates with powerful MCP servers for enhanced capabilities:
flowchart TB
subgraph PAL["PAL MCP - 11 Collaborative Tools"]
P1["chat"] & P2["thinkdeep"] & P3["debug"]
P4["planner"] & P5["precommit"]
P6["codereview"] & P7["consensus"] & P8["challenge"]
P9["apilookup"] & P10["listmodels"] & P11["clink"]
end
subgraph Rube["Rube MCP - 500+ App Integrations (11 tools)"]
R1["SEARCH_TOOLS"] & R2["GET_TOOL_SCHEMAS"] & R3["MULTI_EXECUTE_TOOL"]
R4["REMOTE_BASH_TOOL"] & R5["REMOTE_WORKBENCH"]
R6["FIND_RECIPE"] & R7["EXECUTE_RECIPE"] & R8["GET_RECIPE_DETAILS"]
R9["CREATE_UPDATE_RECIPE"] & R10["MANAGE_CONNECTIONS"] & R11["MANAGE_RECIPE_SCHEDULE"]
end
TASK["SuperClaude Task"] --> PAL
TASK --> Rube
style PAL fill:#e3f2fd
style Rube fill:#fce4ecPAL MCP (11 Tools)
Collaborative intelligence for code review, debugging, and multi-model consensus:
| Tool | Purpose | Use Case |
|---|---|---|
mcp__pal__chat |
General collaborative thinking | Brainstorming, second opinions |
mcp__pal__thinkdeep |
Multi-stage investigation | Complex problem analysis |
mcp__pal__planner |
Interactive sequential planning | Project planning with revision |
mcp__pal__consensus |
Multi-model decision making | Architecture decisions, evaluations |
mcp__pal__codereview |
Systematic code review | Quality assessment, security review |
mcp__pal__precommit |
Git change validation | Pre-commit checks, change impact |
mcp__pal__debug |
Root cause analysis | Complex debugging, issue diagnosis |
mcp__pal__challenge |
Critical thinking | Push back on assumptions |
mcp__pal__apilookup |
API documentation lookup | Current SDK/API information |
mcp__pal__listmodels |
List available models | Model selection and capabilities |
mcp__pal__clink |
External AI CLI linking | Connect to Gemini, Codex, etc. |
Rube MCP (Composio)
Session-based workflow automation with 500+ app integrations:
sequenceDiagram
participant C as Claude Code
participant R as Rube MCP
participant A as External App (Slack, GitHub, etc.)
C->>R: RUBE_SEARCH_TOOLS<br/>{session: {generate_id: true}, queries: [...]}
R-->>C: session_id + available tools + execution plan
C->>R: RUBE_MULTI_EXECUTE_TOOL<br/>{session_id, tools: [...], memory: {...}}
R->>A: Execute tool (e.g., SLACK_SEND_MESSAGE)
A-->>R: Tool result
R-->>C: Results + updated memory
opt Complex Workflow
C->>R: RUBE_CREATE_PLAN<br/>{session_id, use_case, difficulty}
R-->>C: Detailed execution plan
end
opt Recipe Execution
C->>R: RUBE_FIND_RECIPE<br/>{query: "..."}
R-->>C: Matching recipes with recipe_id
C->>R: RUBE_EXECUTE_RECIPE<br/>{recipe_id, input_data}
R-->>C: Recipe execution result
endConnected Services:
| Category | Examples |
|---|---|
| Development | GitHub, GitLab, Bitbucket, Linear, Jira |
| Communication | Slack, Discord, Teams, Email |
| Productivity | Notion, Asana, Trello, Airtable |
| Google Workspace | Gmail, Calendar, Drive, Sheets, Docs |
| Microsoft | Outlook, Teams, OneDrive, SharePoint |
| Social | X (Twitter), LinkedIn, Meta apps |
| AI Tools | Various AI services and APIs |
Web Search (via Rube MCP)
Web search capabilities are available through Rube MCP's LINKUP_SEARCH tool:
- Deep or standard search depth
- Sourced answers with citations and URLs
- Accessed via
LINKUP_SEARCHtool inRUBE_MULTI_EXECUTE_TOOL
Skills System
SuperClaude includes 43 active Claude Code skills in .claude/skills/:
flowchart TB
subgraph Skills[".claude/skills/ (43 active)"]
subgraph Agent["Agent Skills (8)"]
A1["agent-data-engineer"]
A2["agent-fullstack-developer"]
A3["agent-kubernetes-specialist"]
A4["agent-ml-engineer"]
A5["agent-performance-engineer"]
A6["agent-react-specialist"]
A7["agent-security-engineer"]
A8["agent-technical-writer"]
end
subgraph Command["Command Skills (26)"]
C1["sc-analyze"]
C2["sc-brainstorm"]
C3["sc-build"]
C4["sc-cicd-setup"]
C5["sc-code-review"]
C6["sc-design"]
C7["sc-document"]
C8["sc-e2e"]
C9["sc-estimate"]
C10["sc-evaluate"]
C11["sc-explain"]
C12["sc-git"]
C13["sc-implement"]
C14["sc-improve"]
C15["sc-log-fix"]
C16["sc-mcp"]
C17["sc-pr-check"]
C18["sc-pr-fix"]
C19["sc-principles"]
C20["sc-push"]
C21["sc-readme"]
C22["sc-research"]
C23["sc-tdd"]
C24["sc-test"]
C25["sc-workflow"]
C26["sc-worktree"]
end
subgraph GitNexus["GitNexus Skills (6)"]
G1["gitnexus-cli"]
G2["gitnexus-exploring"]
G3["gitnexus-guide"]
G4["gitnexus-impact-analysis"]
G5["gitnexus-debugging"]
G6["gitnexus-refactoring"]
end
subgraph Utility["Utility Skills (3)"]
U1["ask<br/>(single-select)"]
U2["ask-multi<br/>(multi-select)"]
U3["learned<br/>(auto-learned)"]
end
end
subgraph Structure["Skill File Structure"]
direction LR
SKILL["SKILL.md<br/>(AI-facing interface)"]
SCRIPTS["scripts/<br/>(Tool implementations)"]
end
Skills --> Structure
subgraph ScriptDetail["sc-implement/scripts/ (5 tools)"]
S1["select_agent.py (551 lines)<br/>Weighted agent selection"]
S2["run_tests.py (344 lines)<br/>Test framework detection"]
S3["evidence_gate.py (256 lines)<br/>Quality validation"]
S4["skill_learn.py (466 lines)<br/>Skill extraction"]
S5["loop_entry.py (226 lines)<br/>Loop orchestration"]
endSkill Types
| Type | Pattern | Count | Purpose |
|---|---|---|---|
| Agent Skills | agent-* |
8 | Specialized personas for domains |
| Command Skills | sc-* |
26 | Structured workflow implementations |
| GitNexus Skills | gitnexus-* |
6 | Code intelligence, impact analysis, refactoring |
| Utility Skills | ask, ask-multi, learned |
3 | User interaction and learning |
Skill Architecture
Each skill follows a standard structure:
.claude/skills/agent-security-engineer/
└── SKILL.md # AI-facing interface (what Claude reads)
.claude/skills/sc-implement/
├── SKILL.md # Command interface
└── scripts/ # Optional tool implementations
├── select_agent.py # Weighted agent selection algorithm
├── run_tests.py # Test framework detection and execution
├── evidence_gate.py # Quality validation (production_ready/acceptable/needs_review)
├── skill_learn.py # Skill extraction from successful executions
└── loop_entry.py # Loop orchestration entry pointImportant Distinction:
SKILL.md= The interface Claude reads to understand when/how to invokescripts/*.py= Backend implementations executed by the host system
This follows standard agentic patterns: config-first architecture means the AI-facing interface is pure configuration, while Python handles orchestration.
Skill Learning System
The learned/ skill directory stores patterns extracted from successful executions:
- Extraction:
skill_learn.pyextracts patterns from successful--loopexecutions - Promotion: Skills promoted when quality score >= 85.0 and success rate >= 70%
- Retrieval: Auto-retrieved based on task keywords, file types, domain context
- Metadata: Tracks provenance (session ID, timestamp, quality progression)
Using Skills
Skills are invoked via:
- Semantic selection - Claude automatically selects based on task context
- Direct reference -
@agent-security-engineeror/sc-implement - Skill tool -
Skill("agent-security-engineer")for explicit invocation
Quality System
SuperClaude uses a 9-dimension quality scoring system with automated actions:
flowchart LR
subgraph Dimensions["9 Quality Dimensions"]
C["Correctness<br/>25%"]
CO["Completeness<br/>20%"]
P["Performance<br/>10%"]
M["Maintainability<br/>10%"]
S["Security<br/>10%"]
SC["Scalability<br/>10%"]
T["Testability<br/>10%"]
U["Usability<br/>5%"]
PAL["PAL Review<br/>0%*"]
end
subgraph Bands["Quality Bands"]
E["90+ Excellent"]
G["70-89 Good"]
A["50-69 Acceptable"]
P2["30-49 Poor"]
F["0-29 Failing"]
end
subgraph Actions["Auto-Actions"]
FT["Fast-track<br/>Skip review"]
AP["Approve<br/>Complete task"]
IT["Iterate<br/>Loop continues"]
ES["Escalate<br/>Human review"]
BL["Block<br/>Reject output"]
end
Dimensions -->|"Weighted<br/>Average"| Bands
E --> FT
G --> AP
A --> IT
P2 --> ES
F --> BL
style E fill:#4caf50,color:#fff
style G fill:#8bc34a
style A fill:#ffeb3b
style P2 fill:#ff9800,color:#fff
style F fill:#f44336,color:#fff*PAL Review: Dynamic weight when --pal-review enabled
Quality Dimensions
| Dimension | Weight | Indicators |
|---|---|---|
| Correctness | 25% | Tests pass, no runtime errors, output validation |
| Completeness | 20% | Feature coverage, edge cases, documentation |
| Performance | 10% | Time/space complexity, resource usage |
| Maintainability | 10% | Readability, modularity, naming conventions |
| Security | 10% | Input validation, authentication, data protection |
| Scalability | 10% | Architecture patterns, database design, caching |
| Testability | 10% | Unit tests, integration tests, test quality |
| Usability | 5% | UI consistency, error messages, accessibility |
| PAL Review | 0%* | External code review score (dynamic weight) |
Evidence Requirements
| Requires Evidence | Does Not Require Evidence |
|---|---|
sc-implement - File diffs, test results |
sc-analyze - Analysis output |
sc-improve - Before/after, metrics |
sc-design - Specifications, diagrams |
sc-test - Test output, coverage |
sc-document - Documentation files |
sc-git - Git history (self-documenting) |
|
sc-explain - Learning materials |
Quality Thresholds (from evidence_gate.py)
| Status | Score | Action |
|---|---|---|
| Production Ready | 90.0+ | Fast-track, skip additional review |
| Acceptable | 70.0-89.9 | Approve, task complete |
| Needs Review | 50.0-69.9 | Iterate or escalate |
| Below Threshold | < 50.0 | Block, requires significant rework |
Quality Gates
SuperClaude enforces code quality through automated validators integrated into the development workflow via the /sc:principles command.
KISS Validator
Enforces code simplicity via AST analysis. Located at .claude/skills/sc-principles/scripts/validate_kiss.py.
| Metric | Threshold | Severity | Description |
|---|---|---|---|
| Cyclomatic Complexity | > 10 | error | Number of independent paths through code |
| Cyclomatic Complexity | > 7 | warning | Early warning for growing complexity |
| Cognitive Complexity | > 15 | error | Weighted complexity (nested structures count more) |
| Function Length | > 50 lines | error | Lines of code per function |
| Nesting Depth | > 4 levels | error | If/for/while/with/try nesting |
| Parameter Count | > 5 | warning | Function parameters |
Usage:
python .claude/skills/sc-principles/scripts/validate_kiss.py --scope-root . --jsonCognitive vs Cyclomatic Complexity:
- Cyclomatic counts decision points (branches)
- Cognitive weights nested structures more heavily:
1 + nesting_depthper control structure - Example:
if (if (if ...))has low cyclomatic but high cognitive (hard to read)
Purity Validator
Enforces "Functional Core, Imperative Shell" architectural pattern. Located at .claude/skills/sc-principles/scripts/validate_purity.py.
| Layer | Path Patterns | I/O Allowed | Severity |
|---|---|---|---|
| Core (purity required) | */domain/*, */logic/*, */services/*, */utils/*, */core/* |
NO | error |
| Shell (I/O allowed) | */handlers/*, */adapters/*, */api/*, */cli/*, */scripts/*, */tests/* |
YES | warning |
Detected I/O Patterns:
| Category | Examples |
|---|---|
| File I/O | open(), read(), write(), Path.read_text() |
| Network | requests.get(), httpx, urllib, socket |
| Database | execute(), query(), session.add(), cursor |
| Subprocess | subprocess.run(), os.system(), Popen |
| Global State | global, nonlocal keywords |
| Side Effects | print(), logging.*, logger.* |
| Async I/O | async def, await, async for, async with |
Usage:
python .claude/skills/sc-principles/scripts/validate_purity.py --scope-root . --jsonSOLID Validator
Enforces SOLID design principles via AST analysis. Located at .claude/skills/sc-principles/scripts/validate_solid.py.
| Principle | Check | Threshold | Severity |
|---|---|---|---|
| SRP (Single Responsibility) | File length | > 300 lines | error |
| SRP (Single Responsibility) | Class public methods | > 5 | error |
| OCP (Open/Closed) | isinstance cascades | > 2 chained | error |
| LSP (Liskov Substitution) | NotImplementedError in overrides | Any | error |
| ISP (Interface Segregation) | Fat interfaces/protocols | > 7 methods | error |
| DIP (Dependency Inversion) | Direct instantiation in business logic | Any | warning |
Usage:
python .claude/skills/sc-principles/scripts/validate_solid.py --scope-root . --jsonLet It Crash Validator
Enforces the "Let It Crash" error handling philosophy by detecting anti-patterns. Located at .claude/skills/sc-principles/scripts/validate_crash.py.
| Anti-Pattern | Description | Severity |
|---|---|---|
Bare except: |
Catches all exceptions without specificity | error |
except Exception without re-raise |
Swallows exceptions silently | error |
except: pass |
Silent failure hiding bugs | error |
| Nested try/except cascades | Deeply nested error handling (> 2 levels) | error |
Usage:
python .claude/skills/sc-principles/scripts/validate_crash.py --scope-root . --jsonPre-commit Integration
Add to .pre-commit-config.yaml:
repos:
- repo: local
hooks:
- id: kiss-check
name: KISS Validation
entry: python .claude/skills/sc-principles/scripts/validate_kiss.py --scope-root . --json
language: python
types: [python]
pass_filenames: false
- id: purity-check
name: Purity Validation
entry: python .claude/skills/sc-principles/scripts/validate_purity.py --scope-root . --json
language: python
types: [python]
pass_filenames: false
- id: solid-check
name: SOLID Validation
entry: python .claude/skills/sc-principles/scripts/validate_solid.py --scope-root . --json
language: python
types: [python]
pass_filenames: false
- id: crash-check
name: Let It Crash Validation
entry: python .claude/skills/sc-principles/scripts/validate_crash.py --scope-root . --json
language: python
types: [python]
pass_filenames: falseExit Codes
| Code | Meaning | Action |
|---|---|---|
| 0 | Validation passed | Proceed |
| 2 | Violations detected | Blocked - refactor required |
| 3 | Validation error | Manual intervention needed |
Refactoring Guidance
When violations are detected:
For Complexity Violations:
- Extract Method - Split large functions into smaller, named pieces
- Guard Clauses - Replace nested if/else with early returns
- Strategy Pattern - Replace complex switch/if-else with polymorphism
- Decompose Conditional - Name complex conditions as explaining variables
For Purity Violations:
- Dependency Injection - Pass dependencies as arguments
- Repository Pattern - Isolate database operations
- Adapter Pattern - Wrap external APIs
- Return Don't Print - Return values, let callers handle output
Metrics System
SuperClaude includes a callback-based metrics system (core/metrics.py) that decouples the orchestrator from specific metrics backends.
MetricsEmitter Protocol
@runtime_checkable
class MetricsEmitter(Protocol):
"""Protocol defining the interface for metrics emission."""
def __call__(
self,
metric_name: str,
value: Any,
tags: Optional[Dict[str, str]] = None,
) -> None: ...Available Emitters
| Emitter | Purpose | Use Case |
|---|---|---|
noop_emitter |
Does nothing | Default when no emitter configured |
InMemoryMetricsCollector |
Collects to list | Testing and debugging |
LoggingMetricsEmitter |
Logs via Python logging | Simple production logging |
InMemoryMetricsCollector
Testing utility for collecting and querying metrics:
from core.metrics import InMemoryMetricsCollector
collector = InMemoryMetricsCollector()
orchestrator = LoopOrchestrator(config, metrics_emitter=collector)
result = orchestrator.run(context, invoker)
# Query collected metrics
assert collector.get("loop.completed.count") == 1
assert collector.get("loop.duration.seconds") > 0
assert collector.count("loop.iteration.quality_score.gauge") == 3
# Filter by tags
error_metrics = collector.filter_by_tags(
"loop.errors.count",
{"reason": "skill_invocation"}
)Emitted Metrics
Loop Orchestrator Metrics:
| Metric | Type | Description |
|---|---|---|
loop.started.count |
counter | Loop initiated |
loop.completed.count |
counter | Loop finished (tags: termination_reason) |
loop.duration.seconds |
timing | Total loop time |
loop.iterations.total.gauge |
gauge | Iterations executed |
loop.quality_score.final.gauge |
gauge | Final quality score |
loop.errors.count |
counter | Errors (tags: reason) |
loop.iteration.duration.seconds |
timing | Per-iteration timing |
loop.iteration.quality_score.gauge |
gauge | Per-iteration quality |
loop.iteration.quality_delta.gauge |
gauge | Quality change per iteration |
Skill Learning Metrics:
| Metric | Type | Description |
|---|---|---|
learning.skills.applied.count |
counter | Skills injected at start |
learning.skills.extracted.count |
counter | Skills learned (tags: domain, success) |
learning.skills.promoted.count |
counter | Skills auto-promoted (tags: reason) |
Naming Convention
<component>.<subject>.<unit>| Suffix | Type | Use Case |
|---|---|---|
.count |
counter | Events (incremental) |
.gauge |
gauge | Current state (point-in-time) |
.seconds |
timing | Duration measurement |
Integration Examples
Prometheus:
from prometheus_client import Counter, Gauge
counters, gauges = {}, {}
def prometheus_emitter(name, value, tags=None):
labels = tags or {}
if name.endswith('.count'):
if name not in counters:
counters[name] = Counter(name.replace('.', '_'), '', list(labels.keys()))
counters[name].labels(**labels).inc(value)
elif name.endswith('.gauge'):
if name not in gauges:
gauges[name] = Gauge(name.replace('.', '_'), '', list(labels.keys()))
gauges[name].labels(**labels).set(value)StatsD:
import statsd
client = statsd.StatsClient()
def statsd_emitter(name, value, tags=None):
if name.endswith('.count'):
client.incr(name, value)
elif name.endswith('.gauge'):
client.gauge(name, value)
elif name.endswith('.seconds'):
client.timing(name, value * 1000) # Convert to msType System
Core types defined in core/types.py:
TerminationReason
class TerminationReason(Enum):
"""Reasons for loop termination."""
QUALITY_MET = "quality_threshold_met" # Score >= 70
MAX_ITERATIONS = "max_iterations_reached" # Hard cap at 5
ERROR = "improver_error" # Skill execution failure
HUMAN_ESCALATION = "requires_human_review" # Needs human input
TIMEOUT = "timeout" # Wall-clock exceededLoopConfig
@dataclass
class LoopConfig:
"""Configuration for the agentic loop."""
max_iterations: int = 3 # Requested max (user-configurable)
hard_max_iterations: int = 5 # P0 SAFETY: Cannot be overridden
quality_threshold: float = 70.0 # Target score to meet
timeout_seconds: Optional[float] # Wall-clock timeout (optional)
pal_review_enabled: bool = True # Enable PAL MCP review within loop
pal_model: str = "gpt-5" # Model to use for PAL reviewsQualityAssessment
@dataclass
class QualityAssessment:
"""Result of quality assessment for an iteration."""
overall_score: float # Numeric quality score (0-100)
passed: bool # Whether quality threshold was met
threshold: float = 70.0 # The threshold checked against
improvements_needed: list[str] # Specific improvements to make
metrics: dict[str, float] # Detailed breakdown by dimension
band: str = "unknown" # Quality band (excellent/good/acceptable/poor)
metadata: dict[str, Any] # Additional context from evidenceIterationResult
@dataclass
class IterationResult:
"""Result of a single loop iteration."""
iteration: int # Zero-based iteration number
input_quality: float # Quality score before this iteration
output_quality: float # Quality score after this iteration
improvements_applied: list[str] # Improvements that were applied
time_taken: float = 0.0 # Duration in seconds
success: bool = False # Whether iteration improved quality
termination_reason: str = "" # If loop terminated, why
pal_review: Optional[Dict] = None # PAL MCP review results
changed_files: list[str] # Files modified in this iterationLoopResult
@dataclass
class LoopResult:
"""Final result of complete loop execution."""
final_output: dict[str, Any] # Final state after all iterations
final_assessment: QualityAssessment # Quality assessment of final state
iteration_history: list[IterationResult] # All iteration results
termination_reason: TerminationReason # Why the loop stopped
total_iterations: int # Number of iterations executed
total_time: float = 0.0 # Total wall-clock timeConfiguration
SuperClaude uses 6 YAML configuration files:
config/superclaud.yaml
Main framework configuration:
version: 7.0.0
name: SuperClaude Framework
modes:
default: normal
available: [normal, brainstorming, introspection, task_management, token_efficiency, orchestration]
agents:
max_delegation_depth: 5
circular_detection: true
parallel_execution: true
default_timeout: 300
commands:
prefix: "/sc:"
discovery_path: SuperClaude/Commands
cache_ttl: 3600
quality:
enabled: true
default_threshold: 70.0
max_iterations: 5
dimensions:
correctness: 0.25
completeness: 0.20
performance: 0.10
maintainability: 0.10
security: 0.10
scalability: 0.10
testability: 0.10
usability: 0.05
pal_review: 0.00 # Dynamic when enabled
mcp_servers:
enabled: true
timeout: 300
retry_attempts: 3
servers:
pal:
enabled: true
models: [gpt-5, claude-opus-4.1, gemini-2.5-pro]
triggers:
mode_triggers:
brainstorming: [explore, brainstorm, "figure out", "not sure"]
introspection: ["analyze my reasoning", reflect, meta-cognitive]
agent_triggers:
communicator: [debug, error, broken, "not working", explain]
optimizer: [refactor, "improve code", "clean up", optimize]
workflows:
default_workflow: [analyze, plan, implement, test, document]
debug_workflow: [reproduce, analyze, isolate, fix, verify]Quality dimensions, model routing, and consensus policies are configured at runtime:
- Quality dimensions and weights are in
config/superclaud.yaml(quality section) - Model routing is handled by PAL MCP (
mcp__pal__listmodelsat runtime) - Agent selection is filesystem-based (
agents/core/,agents/traits/,agents/extensions/)
Directory Structure
SuperClaude/
├── CLAUDE.md # Master system prompt
├── README.md # This file
├── CONTRIBUTING.md # Contribution guidelines
├── CHANGELOG.md # Version history
├── SONDERA_INTEGRATION.md # Sondera security layer integration guide
├── pyproject.toml # Python project config
├── install-with-sondera.sh # Automated Sondera installer
├── test-sondera-integration.sh # Sondera integration verification
│
├── .claude/
│ └── skills/ # 43 Claude Code skills
│ ├── agent-*/ # 8 agent persona skills
│ │ └── SKILL.md
│ ├── sc-*/ # 26 command skills
│ │ ├── SKILL.md
│ │ └── scripts/ # Optional tool implementations
│ ├── gitnexus/ # 6 GitNexus code intelligence skills
│ │ ├── gitnexus-cli/
│ │ ├── gitnexus-exploring/
│ │ ├── gitnexus-guide/
│ │ ├── gitnexus-impact-analysis/
│ │ ├── gitnexus-debugging/
│ │ └── gitnexus-refactoring/
│ ├── ask/ # Single-select questions
│ ├── ask-multi/ # Multi-select questions
│ └── learned/ # Auto-learned skills
│
├── agents/
│ ├── index.yaml # Agent registry
│ ├── core/ # 5 core agent prompts
│ │ ├── architect.md
│ │ ├── developer.md
│ │ ├── optimizer.md
│ │ ├── guardian.md
│ │ └── communicator.md
│ ├── traits/ # 7 composable trait prompts
│ │ ├── security-first.md
│ │ ├── performance-first.md
│ │ ├── test-driven.md
│ │ ├── minimal-changes.md
│ │ ├── principles-enforced.md
│ │ ├── mcp-pal-enabled.md
│ │ └── mcp-rube-enabled.md
│ ├── extensions/ # 7 domain specialist prompts
│ │ ├── typescript-react-expert.md
│ │ ├── python-expert.md
│ │ ├── golang-expert.md
│ │ ├── rust-expert.md
│ │ ├── kubernetes-specialist.md
│ │ ├── data-engineer.md
│ │ └── ml-engineer.md
│ └── DEPRECATED/ # Archived agents from previous versions (25 files)
│
├── commands/
│ ├── index.yaml # Command registry
│ ├── index.md # Command reference
│ ├── analyze.md
│ ├── implement.md
│ ├── test.md
│ └── ... # 16 command templates
│
├── config/
│ ├── superclaud.yaml # Main framework config (read by Rust crates)
│ └── schemas/ # JSON schemas for validation
│
├── core/ # Python orchestration layer
│ ├── __init__.py
│ ├── loop_orchestrator.py # Loop management (~410 lines)
│ ├── quality_assessment.py # Quality scoring
│ ├── pal_integration.py # PAL MCP signals
│ ├── metrics.py # Callback-based metrics protocol
│ ├── types.py # Core types (150 lines)
│ ├── skill_learning_integration.py
│ └── skill_persistence.py
│
├── SuperClaude/
│ └── Orchestrator/ # SDK-based orchestration
│ ├── __init__.py
│ ├── evidence.py # Evidence collection
│ ├── events_hooks.py # Event hook integration
│ ├── hooks.py # SDK hook integration
│ ├── loop_runner.py # Agentic loop runner
│ ├── quality.py # Quality assessment
│ └── obsidian_hooks.py # Obsidian vault sync hooks
│
├── crates/ # Rust workspace
│ ├── dashboard/ # Tauri v2 desktop dashboard (Leptos WASM frontend)
│ │ └── frontend/src/
│ │ ├── components/ # UI: sidebar, execution_detail, execution_tree, diff_view, ...
│ │ ├── pages/ # inventory, monitor, control, history
│ │ └── state/ # Reactive signals (AppState, ExecutionTree, DTOs)
│ ├── proto/ # Protobuf/gRPC service definitions
│ ├── superclaude-core/ # Shared domain types and utilities
│ ├── superclaude-daemon/ # gRPC daemon (port 50051)
│ ├── superclaude-runtime/ # Rust runtime: agent selector, obsidian vault, loop runner, registry
│ ├── hangman-game/ # Hangman game example
│ └── snake-game/ # Snake game example
│
├── setup/ # Installation and setup system
│ ├── cli/ # CLI commands (install, update, backup, clean, uninstall)
│ ├── components/ # Component installers (agents, commands, core, mcp, modes)
│ ├── core/ # Installer base, registry, validator
│ ├── services/ # Service integrations
│ │ ├── obsidian_config.py # Obsidian config management
│ │ ├── obsidian_vault.py # Vault reading/scanning
│ │ ├── obsidian_context.py # Context generation
│ │ ├── obsidian_artifact.py # Decision artifact writing
│ │ └── obsidian_md.py # CLAUDE.md integration
│ └── utils/ # Environment, logging, security, UI, updater
│
├── mcp/ # MCP integration guides
│ ├── MCP_Pal.md # PAL MCP (11 tools)
│ ├── MCP_Rube.md # Rube MCP (500+ apps, including web search)
│ └── MCP_Zen.md # Zen MCP
│
├── tests/ # Test suite
│ ├── core/ # Core module tests
│ ├── github_scripts/ # GitHub Actions script tests
│ ├── integration/ # Integration tests
│ ├── scripts/ # Script unit tests
│ ├── setup/ # Setup/installer tests
│ ├── skills/ # Skill tests
│ ├── orchestrator/ # Orchestrator tests
│ ├── loop/ # Loop tests
│ ├── mcp/ # MCP integration tests
│ ├── e2e/ # End-to-end tests
│ ├── services/ # Service tests
│ └── fixtures/ # Test data
│
├── Docs/ # Documentation
│ ├── claude-code-reference.md # Claude Code reference guide
│ ├── quality-gates.md
│ └── README.md
│
└── archive/ # Archived Python SDK (v5)GitHub Workflows
SuperClaude includes 23 GitHub Actions workflows for CI/CD, AI-powered review, and automation:
| Workflow | File | Trigger | Purpose |
|---|---|---|---|
| CI | ci.yml |
Push/PR | Linting, tests, validation |
| AI Code Review | ai-review.yml |
PR opened | PAL MCP consensus code review |
| PAL Consensus Review | pal-consensus-review.yml |
PR | Multi-model consensus code review |
| Nightly Code Review | nightly-review.yml |
Cron (2 AM UTC) | Proactive quality monitoring with autofix |
| Claude Review Phase 1 | claude-review-phase1.yml |
PR | Initial AI review pass |
| Claude Review Phase 2 | claude-review-phase2.yml |
PR | Deep AI review pass |
| Claude Review Phase 3 | claude-review-phase3.yml |
PR opened | Draft PR creation with AI suggestions |
| AI Issue Triage | ai-issue-triage.yml |
Issue opened | Auto-label and categorize issues |
| Issue to PR | issue-to-pr.yml |
Issue labeled | Auto-generate PR from issue |
| AI Review Cost Monitor | ai-review-cost-monitor.yml |
Cron (6h) / workflow_run | Track AI spend, budget alerts ($50/day, $1000/month) |
| Nightly Docs Update | nightly-docs-update.yml |
Cron (4 AM UTC) | Detect code changes, update docs via Claude CLI |
| Update Architecture | update-architecture.yml |
Push/Manual | Auto-update ARCHITECTURE.md |
| AI Release Notes | release-notes.yml |
Release created | Generate structured release notes |
| Commit Notifications | commit-notifications.yml |
Push to main | Slack notifications via Rube MCP |
| Security | security.yml |
Push/PR | Security scanning |
| Autonomous Code Scanner | autonomous-code-scanner.yml |
Cron/Manual | Proactive code scanning with autofix |
| GitNexus Impact | gitnexus-impact.yml |
PR | PR impact analysis via GitNexus |
| Package Publish | package-publish.yml |
Release/Manual | PyPI package publishing |
| Package Test | package-test.yml |
Push/PR | Package installation and import testing |
| Stochastic Evals | stochastic-evals.yml |
Manual | LLM evaluation runs |
| E2E App Generation | e2e-app-generation.yml |
Manual | End-to-end app generation tests |
| README Quality Check | readme-quality-check.yml |
PR | Documentation quality validation |
| Claude Code Action | claude.yml |
PR | Claude Code integration |
All AI workflows use the dual-provider pattern: AWS Bedrock (primary) with Anthropic API fallback.
Creating Custom Agents
Agent Prompt Format
Create a new agent in agents/extensions/:
---
name: my-custom-agent
description: Brief description of the agent's purpose
category: custom
triggers: [keyword1, keyword2, keyword3]
tools: [Read, Write, Edit, Bash, Grep, Glob]
priority: 5
---
# My Custom Agent
You are an expert in [domain] specializing in [specialty].
## Focus Areas
- Area 1: Description
- Area 2: Description
- Area 3: Description
## Approach
1. **Analysis**: Understand the problem thoroughly
2. **Planning**: Design a solution approach
3. **Implementation**: Execute with quality checks
4. **Validation**: Verify results meet requirements
## Best Practices
- Practice 1
- Practice 2
- Practice 3
## Anti-Patterns
- Avoid X because Y
- Don't do Z without W
## Checklist
- [ ] Requirement 1 verified
- [ ] Requirement 2 verified
- [ ] Tests passing
- [ ] Documentation updatedRegister the Agent
Add to agents/index.yaml:
extended:
custom-category:
description: "Custom agents"
priority: 6
agents:
- name: my-custom-agent
file: extended/custom-category/my-custom-agent.md
triggers: [keyword1, keyword2]
priority: 5Create Matching Skill
Create .claude/skills/agent-my-custom-agent/SKILL.md:
---
name: agent-my-custom-agent
description: Brief description matching the agent
---
# My Custom Agent Skill
[Content matching the agent definition]Contributing
Contributions welcome! Key areas:
Priority Areas
- New Agents: Add specialized agent personas for underserved domains
- Command Templates: Improve command structures and workflows
- MCP Guides: Add integration documentation for new MCP servers
- Configuration: Tune selection algorithms and quality thresholds
- Core Orchestration: Enhance loop mechanics and termination detection
- Skill Learning: Improve pattern extraction and skill promotion
Quality Standards
- All tests pass before completion
90% test coverage for new features
- Python 3.10+ compatibility
- Cross-platform support (Linux, macOS, Windows)
- Comprehensive error handling and logging
- Security best practices for external integrations
Code Style
- Python: Follow PEP 8, use type hints
- Markdown: Use ATX headers, proper nesting
- YAML: 2-space indentation, quotes for strings with special chars
See CONTRIBUTING.md for detailed guidelines.
License
MIT License - see LICENSE for details.
Acknowledgments
- Claude Code team at Anthropic
- MCP server developers (PAL, Rube/Composio)
- All contributors to the SuperClaude framework
SuperClaude v7.0.0
Config-First Meta-Framework for Claude Code
19 Agents (5 Core + 7 Traits + 7 Extensions) | 43 Skills | 27 Commands | 6 Modes | Quality-Driven Loops
Installationen
Sicherheitsprüfung
Quellcode ansehen
tony363/superclaude
Mehr aus dieser Quelle
AI chat subscription
Turn model research into daily AI work.
Use 100+ models, web search, files, and EU-hosted options in one paid chat workspace.
Inference credits
Build with EU-hosted open-source models.
OpenAI-compatible API for GLM, Kimi, DeepSeek and more. Add credits inside the dashboard.
Verwandte Skills
Mehr aus dieser Quelle: tony363/superclaude
So verwenden Sie diesen Skill
Install agent-security-engineer by running npx skills add tony363/superclaude --skill agent-security-engineer in your project directory. Führen Sie den obigen Installationsbefehl in Ihrem Projektverzeichnis aus. Die Skill-Datei wird von GitHub heruntergeladen und in Ihrem Projekt platziert.
Keine Konfiguration erforderlich. Ihr KI-Agent (Claude Code, Cursor, Windsurf usw.) erkennt installierte Skills automatisch und nutzt sie als Kontext bei der Code-Generierung.
Der Skill verbessert das Verständnis Ihres Agenten für agent-security-engineer, und hilft ihm, etablierte Muster zu befolgen, häufige Fehler zu vermeiden und produktionsreifen Code zu erzeugen.
Was Sie erhalten
Skills sind Klartext-Anweisungsdateien — kein ausführbarer Code. Sie kodieren Expertenwissen über Frameworks, Sprachen oder Tools, das Ihr KI-Agent liest, um seine Ausgabe zu verbessern. Das bedeutet null Laufzeit-Overhead, keine Abhängigkeitskonflikte und volle Transparenz: Sie können jede Anweisung vor der Installation lesen und prüfen.
Kompatibilität
Dieser Skill funktioniert mit jedem KI-Coding-Agenten, der das skills.sh-Format unterstützt, einschließlich Claude Code (Anthropic), Cursor, Windsurf, Cline, Aider und anderen Tools, die projektbezogene Kontextdateien lesen. Skills sind auf Transportebene framework-agnostisch — der Inhalt bestimmt, für welche Sprache oder welches Framework er gilt.
AI chat subscription
Turn model research into daily AI work.
Use 100+ models, web search, files, and EU-hosted options in one paid chat workspace.