Smart LLM Agents (SLA)

English · 繁體中文 · 简体中文

Intelligent multi-model development loop for Claude Code — from planning to PR, fully automated.

---

⚡ One-Click Install

curl -fsSL https://raw.githubusercontent.com/fishtvlvoe/sla/main/install.sh | bash

The installer auto-detects your environment and configures everything.

---

What is SLA?

SLA is a Claude Code skill pack that routes tasks to the best AI model automatically. It covers the full development lifecycle: planning → TDD development → code review → PR → CI/CD monitoring — all in one closed loop.

The problem it solves: Different tasks need different models. Planning needs Claude Opus. Code analysis needs Kimi's long context. Quick edits don't need a powerful model at all. SLA handles this routing automatically — you just describe what you want to build.

---

Full Workflow

You: "I want to build X"
         ↓
   /sla:plan  ──── Claude Opus plans the feature
         ↓         claude-in-chrome searches docs if needed
         ↓         Kimi reads existing code for context
   Outputs: PLAN.md + Epics + Stories
         ↓
   /sla:develop ── smart-route assigns per task:
         │          Simple tasks  → Ollama qwen3-coder:cloud
         │          Complex logic → deepseek-v3.1:cloud / Claude Sonnet
         │          Code analysis → Kimi MCP
         │
         │  ┌── TDD Loop ─────────────────────────┐
         │  │  Write test (red) → Implement (green) │
         │  │  → Refactor → Playwright E2E          │
         │  │  → Stuck? → PUA + YES.md kick in      │
         │  └────────────────────────────────────────┘
         ↓
   /sla:review ─── Kimi MCP reads git diff → Code Review report
         ↓         Issues? Back to develop. Clean? Continue.
   /sla:release ── git push → gh pr create → GitHub Actions triggered
         ↓
   /sla:status ─── Monitor CI results
                   ✅ Pass → Loop complete 🎉
                   ❌ Fail → Kimi reads log → Root cause → Back to develop

---

Commands

Development Loop

Command	Description
/smart-route "task"	Auto-route task to best model
/sla:plan "feature"	Plan feature with GSD integration
/sla:develop	TDD development loop
/sla:review	Kimi-powered code review
/sla:release	Push PR + trigger CI
/sla:status	Monitor GitHub Actions

CLAUDE.md Maintenance

Command	Description
AI.MD	Optimize CLAUDE.md format for AI consumption
audit my CLAUDE.md	Audit content accuracy against codebase (official)
/revise-claude-md	Capture session learnings into CLAUDE.md

Self-Improvement

Command	Description
/learn	Auto-record pitfalls in background
/diary	Write session reflection diary
/reflect	Weekly pattern analysis + suggest CLAUDE.md updates

---

Model Routing

Role	Primary	Fallback	Use Case
PLANNER	Claude Opus	deepseek:671b-cloud	Planning, architecture
CODER	Claude Sonnet	qwen3-coder:480b-cloud	Implementation
FAST_CODER	qwen3-coder:480b-cloud	qwen3:8b (local)	Quick changes, CSS
ANALYZER	Kimi MCP	Claude Sonnet	Code analysis, large files
REVIEWER	Kimi MCP	Claude Sonnet	Code review, diff analysis
BROWSER	claude-in-chrome	Playwright	Web search, UI validation
DEBUGGER	Claude Sonnet + PUA	deepseek:cloud	Stuck debugging

---

Requirements

Tool	Required	Auto-installed
Claude Code	✅ Yes	No (assumed)
Ollama	✅ Yes	✅ Yes
GitHub CLI (gh)	✅ Yes	✅ Yes
GSD Framework	Recommended	No (manual)
Kimi API Key	Optional	No (guided)
PUA Skill	Recommended	✅ Yes
YES.md Skill	Recommended	✅ Yes

---

Ollama Cloud Models

SLA uses Ollama's cloud models — no local GPU required.

# Used by SLA automatically
ollama run qwen3-coder:480b-cloud    # Fast coding
ollama run deepseek-v3.1:671b-cloud  # Complex reasoning
ollama run glm-4.6:cloud             # Chinese reasoning

---

Anti-Failure Mechanisms

SLA bundles two complementary skills to keep AI on track:

PUA — Prevents giving up:

Failure × 2 → L1: Mild pressure → Switch approach
Failure × 3 → L2: WebSearch + read source code
Failure × 4 → L3: 7-item systematic checklist
Failure × 5 → L4: All-in mode

YES.md — Enforces correctness:

• Evidence before conclusions (no guessing)
• Backup before modifying files
• Verify after every fix
• Ripple check: look for related issues

PUA keeps you going. YES.md keeps you going correctly.

---

Uninstall

curl -fsSL https://raw.githubusercontent.com/fishtvlvoe/sla/main/uninstall.sh | bash

---

Contributing

PRs welcome! Please read our contributing guide.

---

Credits

• tanweai/pua — PUA debugging skill
• sstklen/yes.md — Engineering discipline skill
• sstklen/ai-md — CLAUDE.md AI-native optimizer
• anthropics/claude-plugins-official — CLAUDE.md management (official)
• GSD Framework — Development workflow
• Kimi MCP — Long-context analysis

License

MIT