gstack-coder

name

gstack-coder

description

Multi-tool coding skill integrating gstack (Claude Code skills), pi (pi-coding-agent), and mini (mini-swe-agent). Use when spawning Claude Code sessions for coding work, running security audits, code reviews, QA testing URLs, building features end-to-end, or planning before building. Also configures and orchestrates pi and mini to maximize utilization of all AI subscriptions and APIs. For MCP-equipped sub-agent dispatch with provider fallback, pair with pi-mini-orchestrator skill.

GStack Coder — Multi-Tool Coding Orchestration

This skill combines gstack (Claude Code skill pack) with pi and mini to form a complete multi-tool coding workflow. Each tool has unique strengths and subscription quotas — this skill ensures all are utilized.

PRD Anchor: Always read llms.txt in the repository root first. It is the authoritative Product Requirements Document for this repo — it defines architecture, security policy (PQC mandates, secrets management), workflow rules (git worktrees, commit conventions), and the provider/MCP infrastructure. If llms.txt and any other doc conflict, llms.txt wins. Never skip reading it.

---

1. GStack Installation & Setup

gstack is a collection of Claude Code skills by garrytan. Install once:

git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup

After installation, gstack provides these slash commands inside Claude Code:

Command	Purpose
/cso	Security audit — full codebase security review
/review	Code review — detailed PR/code review
/qa https://...	QA test a URL — test a deployed endpoint
/autoplan	Auto-plan — generate implementation plan for a feature
/ship	Ship — build, test, and deploy pipeline
/office-hours	Office hours — architect/planning discussion before building

---

2. Multi-Tool Coding Workflow

Tool Inventory

Tool	CLI	Config Location	Best For
Claude Code (gstack)	claude	~/.claude/	Security audits, code review, planning, shipping
pi	pi (wrapper)	~/.pi/agent/	Fast non-interactive coding, extensions, themes, MCP servers
mini	mini (wrapper)	~/.config/mini-swe-agent/.env	Lightweight SWE agent, bash-based tasks

Provider Hot-Swap (use all subscriptions)

The ainish-coder framework supports hot-swapping provider configs via CLI wrappers. Each tool can be pointed at different API providers to maximize quota usage:

# pi with provider — wrapper auto-swaps config before launch
pi modal -p "Add input validation to src/api/handlers.ts" --no-session --thinking off

# mini with provider
mini nvidia --task "Fix the race condition" --yolo

# MCP-equipped dispatch (see pi-mini-orchestrator skill)
pi modal --mcp tavily-search,github -p "Review PR for security issues" --no-session --thinking off

Canonical provider order: modal → nvidia → nebius → opencode → zai → wafer-serverless → openrouter → zenmux

Always dispatch in this order. On failure, advance to the next. Never skip or reorder.

---

3. Coding Task Patterns

Security Audit

Load gstack. Run /cso

Full codebase security review using gstack's security audit skill. Covers PQC compliance, injection vectors, secret leaks, and dependency vulnerabilities.

Code Review

Load gstack. Run /review

Detailed PR/code review with line-by-line analysis, best practices checks, and improvement suggestions.

QA Test a URL

Load gstack. Run /qa https://staging.example.com

Test a deployed URL end-to-end: functional testing, API validation, UI checks, performance baseline.

Build a Feature End-to-End

Load gstack. Run /autoplan, implement the plan, then run /ship

Full cycle: auto-generate implementation plan, execute, build and ship.

Plan Before Building (No Implementation)

Load gstack. Run /office-hours then /autoplan. Save the plan, don't implement.

Architecture discussion first, then detailed plan. Save for later. No code changes.

---

4. OSA: Orchestrated System of Agents

When dispatching work across coding tools using fixed-order rotation:

gemini → claude → crush → pi → mini → opencode → kilo → (wrap)

YOLO Dispatch Commands

# Sequential dispatch — each agent does one task
gemini -p "Add error handling to src/api/routes.ts" -y
claude -p "Review the error handling for edge cases" --dangerously-skip-permissions
pi modal -p "Add unit tests for the error handling" --no-session --thinking off
mini nvidia --task "Validate routes against schema" --yolo

Parallel Dispatch (separate worktrees)

# Each agent works on a different file in isolation
git worktree add ../task-auth agent-gemini
cd ../task-auth && gemini -p "Implement OAuth2 login flow" -y &

git worktree add ../task-payment agent-pi
cd ../task-payment && pi modal -p "Implement Stripe payment integration" --no-session --thinking off &

wait
# Review and merge worktrees back

---

5. Tool-Specific Configuration

pi (pi-coding-agent)

• Config: ~/.pi/agent/settings.json, auth.json, models.json
• Non-interactive: pi <provider> -p "prompt" --no-session --thinking off
• With MCP: pi <provider> --mcp server1,server2 -p "prompt" --no-session --thinking off
• Extensions: pi -e extensions/your-extension.ts

mini (mini-swe-agent)

• Config: ~/.config/mini-swe-agent/.env
• Non-interactive: mini <provider> --task "prompt" --yolo
• With MCP: mini <provider> --mcp server1,server2 --task "prompt" --yolo
• Config override: mini -c custom.yaml --task "prompt" --yolo

Provider switching

• Both wrappers auto-detect provider arg and hot-swap config before launch
• Original config restored when the tool exits
• All providers in ~/.config/ainish-coder/providers.json

---

6. Provider Configuration

Configure providers in ~/.config/ainish-coder/providers.json. The canonical dispatch order is fixed:

modal → nvidia → nebius → opencode → zai → wafer-serverless → openrouter → zenmux

{
  "modal": {
    "baseUrl": "https://modal.com/api/v1",
    "apiKey": "sk-modal-...",
    "defaultModel": "zai-org/GLM-5.1-FP8",
    "tools": { "pi": true, "mini": true }
  },
  "nvidia": {
    "baseUrl": "https://integrate.api.nvidia.com/v1",
    "apiKey": "nvapi-...",
    "defaultModel": "deepseek/deepseek-v4-flash",
    "tools": { "pi": true, "mini": true }
  },
  "nebius": {
    "baseUrl": "https://api.nebius.com/v1",
    "apiKey": "sk-nebius-...",
    "defaultModel": "NousResearch/Hermes-4-405B",
    "tools": { "pi": true, "mini": true }
  }
}

All 8 providers support both pi and mini. The full provider list, models, and MCP server catalog are defined in llms.txt (the authoritative PRD) and ~/.agents/mcp-settings.json.

---

7. MCP Sub-Agent Dispatch

For MCP-equipped sub-agent dispatch with automatic provider fallback, pair with the pi-mini-orchestrator skill. It provides:

• MCP server catalog (10 servers: tavily, context7, brave, github, postgres, filesystem, codewhale, puppeteer, memory, gitnexus)
• Provider fallback chain (modal → nvidia → nebius → opencode → zai → wafer-serverless → openrouter → zenmux)
• Error recovery patterns (provider failure → next provider, MCP failure → retry without)
• Scoped MCP: sub-agents only get the tools they need per task

# Example: code review with GitHub access + provider fallback
pi modal --mcp github -p "Review PR #42" --no-session --thinking off ||
  pi nvidia --mcp github -p "Review PR #42" --no-session --thinking off ||
  pi nebius --mcp github -p "Review PR #42" --no-session --thinking off

---

8. Adding "Coding Tasks" to AGENTS.md

When setting up a new codebase, add this section to AGENTS.md:

## CODING TASKS

When spawning Claude Code sessions for coding work, tell the session to use gstack skills.
Always read llms.txt first — it is the authoritative PRD for this repository.

### Task Patterns

- **Security audit**: "Load gstack. Run /cso"
- **Code review**: "Load gstack. Run /review"
- **QA test a URL**: "Load gstack. Run /qa https://..."
- **Build feature end-to-end**: "Load gstack. Run /autoplan, implement the plan, then run /ship"
- **Plan before building**: "Load gstack. Run /office-hours then /autoplan. Save the plan, don't implement."

### Multi-Tool Dispatch

For parallel work across subscriptions (in canonical order):
- **pi**: `pi <provider> -p "task" --no-session --thinking off`
- **mini**: `mini <provider> --task "task" --yolo`

Provider order: modal → nvidia → nebius → opencode → zai → wafer-serverless → openrouter → zenmux

For MCP-equipped sub-agent dispatch with provider fallback, pair with pi-mini-orchestrator skill.