The one architectural thing that made everything work

Initial version spawned isolated opencode run processes. Each call was stateless: the Executor would invent its own plan instead of following the Architect's output, and the Reviewer had no actual artifacts to inspect. Empty responses were common.

Switching to opencode serve plus opencode run --attach <url> fixed all of it. All three agent calls share the same session state. Context accumulates across the workflow without any manual plumbing.

What the model stack looks like

• OpenCode headless server on port 4096
• qwen3-coder:30b via Ollama (local inference)
• Python + discord.py for the bot
• Target git repo that agents read from and write to

Repo (180 lines of Python): https://github.com/aminrj/agent-forge-bootstrap

Full write-up with architecture diagrams and threat model: https://aminrj.com/posts/building-a-multi-agents-coding-workflow/