Integration Guide

Hybrid Devin-Station Flow (NHA-21)

For complex tasks, use Devin for planning and Mac Studio for implementation to save costs.

1. Plan Phase (Devin):

   • Add the !plan label to a Linear issue.
   • Devin will analyze the task and post an implementation plan as a comment.
   • Devin will also push the plan to docs/plans/{ISSUE_ID}.md and open a draft PR.

2. Implement Phase (Mac Studio):

   • Review Devin's plan.
   • Add the implement label to the Linear issue.
   • The local pipeline will trigger, read the plan from the comment/file, and execute the Coder/Reviewer/Healer agents.

Configuration

The pipeline identifies a Devin-generated plan by:

• The Generated by Devin marker in the plan text.
• The existence of a docs/plans/{ISSUE_ID}.md file.
• Structured sections like ## Files to Change.

Cost Optimization

• Devin handles the high-level reasoning (Plan).
• Local models (Qwen, Llama) handle the heavy lifting (Implement).
• The Architect agent is automatically skipped if a high-quality plan is detected.

LLM providers

This station supports 3 LLM providers with automatic fallback:

Provider	Type	Cost	When used
Studio Ollama	Local (free)	Free	Primary — always preferred
DeepSeek Chat	Cloud API	$0.14/M input	Fallback when Studio offline
Claude Sonnet 4.5	Cloud API	$3/M input	Last resort

Adding a new provider

Edit pipeline.py LLM_CONFIG dict and fallback_llm() function. Add the API key to .env.

Extending the pipeline

The pipeline at pipeline.py is a single file with clear sections:

• Git Operations — clone, branch, commit, push, PR
• LLM — provider selection + fallback
• Telegram — notification sending
• Pipeline Flow — the main CrewAI flow

To add a new stage, insert it in run_pipeline().

Adding a new node

1. Install Tailscale on the new machine
2. Add its Tailscale IP to all .env files
3. If it's a worker: install Docker + Ollama, add a compose profile
4. If it's a client: install OpenCode, point to Studio's Ollama URL

Reusable agents

See AGENTS.md for agent definitions, prompt templates, and skill configurations that work with any AI model (CrewAI, OpenCode, Claude, DeepSeek, etc.).