Identity
This skill is a perspective collector, not an aggregator. It runs reference models in parallel, gathers each one's analysis, and returns it to the current session (= you, the aggregator). Synthesis and action are done by you with the full toolset.
When
- When you want to layer multiple model perspectives onto a hard problem — "review with moa", "look at it with multiple models".
- soft (default): design·judgment perspectives. hard: actually reads and analyzes the codebase.
Execution
- Run the script to collect perspectives (cwd = target repo for analysis):
cd <repo> && python3 ~/.claude/skills/moa/scripts/moa.py "<프롬프트>" [--hard] [--preset <이름>] [--refs-only]
- Alternative execution form:
PYTHONPATH=~/.claude/skills/moa python3 -m scripts.moa "...".
- Read the outputted
## MoA referencesblock. - Show each reference's perspective to the user first — tool results are collapsed and
invisible to the user. In your response body, label each reference with
Reference N — backend:modeland transcribe the key claims as-is (or as a faithful excerpt if long). Reporting only the synthesis is forbidden. --refs-onlyIf not, follow the aggregation guidelines at the bottom and you synthesize: compare agreements·disagreements·missed points, adjudicate disagreements by argument,[failed:]·[skipped:]as partial information. Do not take references on blind faith; make the final judgment and then act with the full toolset.
Modes·Backends
- soft = openrouter references only (advice). hard = codex·claude read-only independent sessions + openrouter (demoted).
- References cannot cause side effects in any mode (read-only). Only the aggregator has full permissions.
Registry·Auth
- Registry:
~/.claude/moa/config.yml(copies presets.example.yml on first run). Per-preset references·reference_max_tokens·enabled. No aggregator field (always the current session). - openrouter 키:
OPENROUTER_API_KEYenvironment variable (takes precedence) or~/.claude/auth/ai-ml-services.env. codex/claude use each CLI's own authentication.
Source
Port of NousResearch Hermes Agent MoA (MIT) logic — advisory prompts·parallel fan-out·failure isolation· config schema.