hemagent-max

Lean bioinformatics AI agent plugin for OpenCode.

Same intelligence as hemagent, 46% fewer tokens per message.

hemagent vs hemagent-max

| | hemagent | hemagent-max | |---|---|---| | Agent prompt | 164 lines (~2,232 tokens) | 42 lines (~500 tokens) | | Domain knowledge per message | ~1,875 tokens (all 3 docs) | ~600 tokens (1 most relevant) | | Total tokens per bio message | ~3,900 | ~2,100 | | Token savings | — | 46% | | Hooks | 8 | 8 (same) | | Skills | hemtools | hemtools (same) | | Tests | 25 | 25 (same) |

Philosophy

hemagent teaches the LLM how to think about bioinformatics via a long prompt. hemagent-max trusts that the LLM already knows bioinformatics and only tells it what to do differently: load skills first, use checklist format, check tools on first message.

The intelligence lives in hooks (intent classification, skill matching, plan tracking, error recovery), not in the prompt.

Install

# For LLM agents:
curl -fsSL https://raw.githubusercontent.com/YichaoOU/hemagentMax/main/docs/guide/installation.md

Manual

cd hemagentMax && bun install && bun run build
cd ~/.config/opencode && npm install hemagent-max

Add to ~/.config/opencode/opencode.jsonc:

{ "plugin": ["hemagent-max"] }

What It Does

Same as hemagent — 8 hooks for intent classification, skill-first enforcement, plan tracking, error recovery, memory, and system prompt injection. The only differences:

1. Agent prompt is 81% smaller — 5 rules instead of 15 sections
2. Injects 1 know-how doc per message instead of all 3
3. First-run: agent checks installed bio tools (which samtools STAR...) on first message

License

MIT