Autoresearch loop — profile optimizer (layouts + animations + computer use)

[thisis-romar]

Epic tracking the full autoresearch optimizer pipeline for Stream Deck profiles.

What ships

• Loop A (done): Claude API iterates on generate-from-shortcuts.js
• Loop B (done): Claude Code subagent iterates on built profiles via MCP tools
• Loop C (planned): Codex CLI + computer use as the agent brain
• Animated icons (planned): GIF icon generation per shortcut category
• Visual eval (planned): screenshot-based scoring via computer use
• Eval harness (planned): branch isolation, locked evals, checkpoints.log per operating protocol

Shared infrastructure

• scripts/autoresearch/score.mjs — composite scorer (coverage 40% + P1-density 35% + coherence 25% + visual 0-15%)
• autoresearch/program.md — research strategy
• scripts/quality-gate.js — locked Tier 2 evaluator

Sub-tasks

• Numeric profile scorer (score.mjs)
• Research loop orchestrator Loop A (loop.mjs)
• Agent research program (program.md)
• npm scripts + devDep wire-up
• Subagent driver (autoresearch-driver.md)
• Orchestration skill (SKILL.md)
• Loop C — Codex CLI brain (loop-c.mjs)
• Animated GIF icon support
• Visual eval via computer use
• Formal eval harness (branch isolation + checkpoints.log)
• Loop C program doc (loop-c-program.md)