calab-plugin

---

The Problem

AI-assisted development suffers from inconsistency: varying code quality, lost context between sessions, missing documentation, and unverified outputs. Teams waste time compensating for what the AI should handle automatically.

The Solution

Calab Plugin wraps Claude Code in a structured development lifecycle — every plan gets reviewed, every implementation gets tested, every output gets validated. Context persists across sessions. Quality gates are non-negotiable.

Without Calab	With Calab
Code quality varies by prompt	Enforced quality gates on every output
Context lost on compact/restart	Automatic state persistence & restoration
Documentation drifts from code	Auto-synchronized on every change
Hallucinated code ships unchecked	Validator → Reinforcer verification chain
Ad-hoc development process	Plan → Design → Tasks → Build pipeline
No traceability across phases	Mandatory artifacts with structured handoffs

---

Architecture

graph TD
    A[User Request] --> B["/dev — Pipeline"]
    A --> C["/solve — Debugger"]
    A --> D["/onboard — Analyzer"]

    B --> B1[Plan]
    B --> B2[Discuss]
    B --> B3[Design]
    B --> B4[Tasks]
    B --> B5["Build (TDD)"]

    C --> B5

    B5 --> I{Validator}
    I -->|"90%+ ✓"| J[Ship]
    I -->|"70-89%"| K[Reinforcer]
    K --> I
    I -->|"< 70%"| L[Escalate]

    style A fill:#f8fafc,stroke:#334155
    style B fill:#dbeafe,stroke:#2563eb
    style C fill:#fef3c7,stroke:#d97706
    style D fill:#d1fae5,stroke:#059669
    style J fill:#d1fae5,stroke:#059669
    style L fill:#fee2e2,stroke:#dc2626

Skill Layers

Layer	Skills	Trigger
Core	/dev · /solve · /onboard	User-invoked
Utility	/docs · /security · /research · /jira · /refactor · /e2e · /guard	User-invoked
Passive	best-practices · code-quality · tdd-workflow · project-rules · work-tracker · clarification-protocol · skill-completion-rules · verify-agents	Auto-loaded contextually

Lifecycle Hooks (27)

Session start/stop, pre-compact state save, code quality validation, worktree tracking, build error detection, artifact verification, confidence-based reinforcer escalation, and more — all running automatically in the background.

---

Key Features (v2.9.0)

Prompt Engineering Best Practices

Feature	Description
6-Element Task Specification	Every task includes What/How/Avoid+WHY/Verify/Done/Files with specificity testing
Structured Returns	All agents communicate via fixed JSON schemas for reliable inter-agent handoffs
Deviation Rules	Auto-fix protocol for 5 safe categories; user confirmation for 4 high-risk categories
50% Context Budget Rule	4-tier quality management (PEAK/GOOD/DEGRADING/POOR) based on context usage
Checkpoint Classification	human-verify (90%) · decision (9%) · human-action (1%) — minimizes unnecessary interruptions
3-Level Artifact Verification	Existence → Substantive → Wired — ensures artifacts are real, meaningful, and connected
Goal-Backward Verification	Validates from user goal → observable truth → code artifact → key link
Fresh Context Pattern	Each executor starts with only its task definition — no cross-contamination
Wave-based Parallel Execution	Dependency graph → topological sort → wave grouping for parallel task execution
Questioning Guide	"Think partner, not interviewer" — Ask vs Decide framework to minimize user fatigue

Roadmap & Phase Management

/dev --roadmap                       # View roadmap status
/dev --roadmap add "Phase title"     # Add phase
/dev --roadmap insert N "Title"      # Insert urgent phase before N
/dev --roadmap complete N            # Complete phase N
/dev --roadmap milestone "v1.0.0"    # Create milestone + archive

---

Quick Start

# 1. Add marketplace
/plugin marketplace add Wondermove-Inc/calab-claude-plugin

# 2. Install plugin
/plugin install calab-plugin@calab-marketplace

# 3. Verify installation
/plugins

# 4. Onboard your project
/onboard