Hoangsa
HOANGSA is a context engineering system for Claude Code. It solves a fundamental problem: Claude's output quality degrades as the context window fills up.
Install / Use
/learn @studium-ignotum/HoangsaQuality Score
Category
Development & EngineeringSupported Platforms
README
HOANGSA
A context engineering system for Claude Code — split work into bounded tasks, each with a fresh context window.
What is HOANGSA?
HOANGSA is a context engineering system for Claude Code. It solves a fundamental problem: Claude's output quality degrades as the context window fills up.
The fix is structural. HOANGSA splits work into discrete tasks. Each task runs in a fresh context window with only the files it actually needs. The result is consistent, high-quality output across arbitrarily large projects.
The core pipeline:
| Phase | Command | Output |
|-------|---------|--------|
| Design | /hoangsa:menu | DESIGN-SPEC + TEST-SPEC |
| Plan | /hoangsa:prepare | Executable task DAG (plan.json) |
| Execute | /hoangsa:cook | Working code, wave by wave |
| Test | /hoangsa:taste | Acceptance test results |
| Commit | /hoangsa:plate | Conventional commit |
| Review | /hoangsa:ship | Code + security review, push/PR |
The orchestrator never writes code. It dispatches workers, each with a bounded context, and assembles results.
Features
Context Engineering — Each worker task runs in a fresh context window (200k tokens). The plan's context_pointers tell each worker exactly which files to read — no more, no less.
Spec-Driven Development — Every feature starts with a DESIGN-SPEC and TEST-SPEC. Workers implement against specs, not vague instructions. Adaptive spec format for different task types (code, ops, infra, docs).
DAG-Based Execution — Tasks organized as a directed acyclic graph. Independent tasks execute in parallel waves, dependent tasks execute sequentially. No unnecessary serialization.
3-Tier Verification — Each task goes through static analysis, behavioral tests (x3), and semantic review against spec before proceeding.
Cross-Layer Bug Tracing — /hoangsa:fix traces bugs across FE/BE/API/DB boundaries to find the real root cause before touching any code.
Pre-Ship Review Gates — /hoangsa:ship runs code quality and security reviews in parallel, blocks on critical issues, and handles push or PR creation.
8-Dimension Codebase Audit — /hoangsa:audit scans for code smells, security vulnerabilities, performance bottlenecks, tech debt, test coverage gaps, dependency risks, architectural violations, and documentation gaps.
Task Manager Integration — Bidirectional sync with ClickUp and Asana. Pull task details as context, push status/comments/reports back after work completes.
GitNexus Code Intelligence — Built-in call graph analysis. Impact analysis before edits, safe renames across the codebase, and full execution flow tracing.
Visual Debugging — Analyze screenshots and screen recordings. Extract frames from video, generate montages, and overlay diffs to spot visual regressions.
Git Flow Management — Built-in skill for task branching: start, switch, park, resume, finish, cleanup, sync. Auto-detects branching strategy and naming conventions.
Framework-Specific Worker Rules — 15 framework addons (React, Next.js, Vue, Svelte, Angular, Express, NestJS, Go, Rust, Python, Java, Swift, Flutter, TypeScript, JavaScript) tune worker behavior per tech stack.
Multi-Profile Model Selection — Switch between quality, balanced, and budget model profiles to match task requirements and cost constraints.
Quick Start
npx hoangsa-cc # Install HOANGSA into your Claude Code environment
/hoangsa:init # Initialize project — detect codebase, set preferences
/hoangsa:menu # Design your first task
After /hoangsa:menu completes, follow with /hoangsa:prepare to generate a plan, then /hoangsa:cook to execute it.
Installation
Prerequisites: Node.js 18+ and the Claude Code CLI
# Interactive — asks whether to install globally or locally
npx hoangsa-cc
# Install to ~/.claude/ — available in all projects
npx hoangsa-cc --global
# Install to .claude/ — this project only
npx hoangsa-cc --local
# Remove HOANGSA
npx hoangsa-cc --uninstall
| Flag | Short | Description |
|------|-------|-------------|
| --global | -g | Install to ~/.claude/ (all projects) |
| --local | -l | Install to .claude/ (this project only) |
| --uninstall | -u | Remove HOANGSA |
The installer also sets up:
- Lifecycle hooks (statusline, context monitor, update checker, GitNexus tracker)
- GitNexus MCP server for code intelligence
- Task manager MCP integration (if configured)
- Quality gate skills (silent-failure-hunter, pr-test-analyzer, comment-analyzer, type-design-analyzer)
Workflow
idea → /menu Design → DESIGN-SPEC + TEST-SPEC
→ /prepare Plan → Executable task DAG (plan.json)
→ /cook Execute → Wave-by-wave, fresh context per task
→ /taste Test → Acceptance tests per task
→ /plate Commit → Conventional commit message
→ /ship Review → Code + security gates, push/PR
→ /serve Sync → Bidirectional task manager sync
Design (/menu) — Interview the user about requirements. Produce a structured DESIGN-SPEC with interfaces and acceptance criteria, plus a TEST-SPEC with test cases and coverage targets.
Plan (/prepare) — Parse the specs and generate plan.json: a DAG of tasks, each with an assigned worker, bounded file list (context_pointers), and explicit dependency edges.
Execute (/cook) — Walk the DAG wave by wave. Dispatch each worker with its context. Independent tasks in the same wave run in parallel. Each completed task goes through an auto-simplify pass before advancing.
Test (/taste) — Run the acceptance tests defined in TEST-SPEC. Report pass/fail per task. Block the pipeline on failures, delegate fixes to /hoangsa:fix.
Commit (/plate) — Stage changes and generate a conventional commit message from the completed work.
Review (/ship) — Launch parallel code quality and security review agents. Block on critical/high issues. User decides: fix, override, or cancel. On pass, push and/or create PR with review summary.
Sync (/serve) — Push status updates, comments, and artifacts back to the linked task manager.
Commands
Core Workflow
| Command | Description |
|---------|-------------|
| /hoangsa:menu | Design — from idea to DESIGN-SPEC + TEST-SPEC |
| /hoangsa:prepare | Plan — convert specs to an executable task DAG |
| /hoangsa:cook | Execute — wave-by-wave with fresh context per task |
| /hoangsa:taste | Test — run acceptance tests per task |
| /hoangsa:plate | Commit — generate and apply a conventional commit message |
| /hoangsa:ship | Ship — code + security review, then push or create PR |
| /hoangsa:serve | Sync — bidirectional sync with connected task manager |
Specialized
| Command | Description |
|---------|-------------|
| /hoangsa:fix | Hotfix — cross-layer root cause tracing + minimal targeted fix |
| /hoangsa:audit | Audit — 8-dimension codebase scan (security, debt, coverage, etc.) |
| /hoangsa:research | Research — codebase analysis combined with external research |
Utility
| Command | Description |
|---------|-------------|
| /hoangsa:init | Initialize — detect codebase, configure preferences, first-time setup |
| /hoangsa:check | Status — show current session progress and pending tasks |
| /hoangsa:index | Index — rebuild GitNexus code intelligence graph |
| /hoangsa:update | Update — upgrade HOANGSA to the latest version |
| /hoangsa:help | Help — show all available commands |
Skills
HOANGSA includes built-in skills that extend Claude Code's capabilities:
Git Flow
Task-oriented git workflow management. Start a task branch, park work-in-progress, switch between tasks, and finish with push + PR — all with dirty-state guards and auto-detection of your branching strategy.
Flows: start | switch | park | resume | finish | cleanup | sync
Visual Debug
Analyze screenshots and screen recordings to debug visual issues. Extracts frames from video files, generates montage grids for overview, and creates diff overlays to highlight changes between frames.
Supports: .png, .jpg, .webp, .gif, .mp4, .mov, .webm, .avi, .mkv
Configuration
HOANGSA stores project configuration in .hoangsa/config.json.
{
"lang": "en",
"spec_lang": "en",
"tech_stack": ["typescript", "react", "postgres"],
"review_style": "strict",
"model_profile": "balanced",
"task_manager": {
"provider": "clickup",
"token": "<your-token>"
}
}
Preferences
| Key | Values | Description |
|-----|--------|-------------|
| lang | en, vi | Language for orchestrator output |
| spec_lang | en, vi | Language for generated specs |
| tech_stack | array | Project technology stack (used to select worker rule addons) |
| review_style | strict, balanced, light | Code review thoroughness |
| interaction_level | minimal, standard, detailed | How much the orchestrator asks |
Model Profiles
Select a profile to control the model used at each role:
| Profile | Worker | Designer | Reviewer |
|---------|--------|----------|----------|
| quality | claude-opus | claude-opus | claude-opus |
| balanced | claude-sonnet | claude-opus | claude-sonnet |
| budget | claude-haiku | claude-sonnet | claude-haiku |
Switch profiles with /hoangsa:init or by editing model_profile in config.json.
Task Manager
Related Skills
node-connect
351.8kDiagnose OpenClaw node connection and pairing failures for Android, iOS, and macOS companion apps
frontend-design
110.9kCreate distinctive, production-grade frontend interfaces with high design quality. Use this skill when the user asks to build web components, pages, or applications. Generates creative, polished code that avoids generic AI aesthetics.
openai-whisper-api
351.8kTranscribe audio via OpenAI Audio Transcriptions API (Whisper).
qqbot-media
351.8kQQBot 富媒体收发能力。使用 <qqmedia> 标签,系统根据文件扩展名自动识别类型(图片/语音/视频/文件)。
