henrijss/opencode-simple-orchestrator

This repository is a personal configuration for the Opencode AI coding tool's orchestrator feature, shared publicly so others can use it as a starting point. The author's README warns clearly that the setup is specific to their own workflow and models, and anyone borrowing it should adapt it rather than copy it directly. The core idea is to run one capable, expensive AI model as a central coordinator that breaks a task into pieces and hands each piece to a cheaper, more focused subagent. The author pairs GPT 5.5 as the main agent with DeepSeek v4 Flash for the subagents. The main agent is permitted to use all tools as a fallback if the subagents fail, but otherwise stays out of direct execution. Three subagents are configured. Scout handles documentation research using Deepwiki, Context7, and web crawlers to look up references. Explore is a read-only agent that verifies files and maps the structure of a codebase. General is the workhorse agent that runs commands and edits files, with access to the full tool set. When a task comes in, the main agent works through five phases: map the relevant parts of the codebase, gather context from those areas, produce a plan, spawn one General subagent per domain to carry out the plan independently, and then spawn verification agents to review the results. The configuration files are intentionally verbose because restricting each agent to only the tools it needs is central to keeping costs low and behavior predictable. The project contains no application code, only prompt and tool configuration intended to be reworked for a different person's setup.