pi-moa

Mixture-of-agents workflow for pi, adapted from Hermes MoA for coding-agent use.

What it adds

• A normal moa model provider with one model per preset.
• A /moa [preset] <prompt> one-shot command that restores your previous model.
• Advisor/reference models run on manual MoA invocations, not every tool-result turn.
• Normal-model turns get guidance to ask before using MoA for unusually complex work.
• Advisors see only user/assistant text, without tool schemas or tool results.
• The aggregator keeps normal pi tools and acts as the real coding agent.
• Advisor outputs are visible as a collapsible thinking block and filtered from future context.
• Global config plus optional project override.

Install

pi install npm:@peeraponw/pi-moa

Or try without installing:

pi -e npm:@peeraponw/pi-moa

You can also install straight from GitHub:

pi install git:github.com/peeraponw/pi-moa

Configure

Run:

/moa setup

The default generated config uses:

• advisors: zai/glm-5.2:xhigh, openai-codex/gpt-5.5:xhigh
• aggregator: zai/glm-5.2:high
• presets: default, architect, bug, review, plan, debug

Config files:

• global: ~/.pi/agent/moa.json
• project override: .pi/moa.json

Reload after editing:

/moa reload

Use

One-shot:

/moa architect design the migration
/moa bug find the race condition
/moa review review my current diff

Persistent model selection:

/model architect --provider moa

With a persistent MoA model, advisors run on user-facing turns and are skipped on follow-up tool-result turns. Their latest advice is reused privately during the same tool loop, so tools stay cheaper without losing the manual MoA guidance.

List presets:

/moa list