‹ 首页

autoresearch

@popmechanic · 收录于 1 周前

Run the Parallel Autoresearch Engine — massively parallel SKILL.md optimization. Generates 10+ SKILL.md variants per generation, tests each with programmatic harness (no browser needed), scores with triple-run averaging, and iterates autonomously. Use when asked to run autoresearch, improve SKILL.md at scale, or start parallel autoresearch.

场景技能开发自动化测试

适合你，如果你需要大规模自动优化和验证 SKILL.md 文件。

/ 下载安装
autoresearch.skill双击,或拖进 Claude 桌面版 / Cowork,即完成安装↓ .skill↓ .zip
用别的 agent?下载 .zip 解压,把文件夹放进它的技能目录
Claude Code~/.claude/skills/(项目级 .claude/skills/)
Codex CLI~/.codex/skills/
Cursor自动读取上面两处目录
其他工具见其文档的「skills」目录;两个下载是同一份文件,只是名字不同
/ 通过 npx 安装 校验哈希
npx oh-my-skill add popmechanic/vibesos/autoresearch
/ 通过 bash 安装
curl -fsSL https://oh-my-skill.com/install.sh | bash -s -- popmechanic/vibesos/autoresearch
/ 已经装过?验证本机副本,不用重装
npx oh-my-skill verify popmechanic/vibesos/autoresearch
安装目标可用 --agent / --scope 或 --to 明确指定;省略时只会在唯一已存在的 agent 目录上自动选择,零命中或多命中会停止并提示。content_hash 缺失或不一致均拒装。
132GitHub stars
~747上下文体积 · 单文件
镜像托管

怎么用

技能原文 SKILL.md作者撰写 · MIT · 13766a0

Parallel Autoresearch Engine

Plan mode: This skill is ONE plan step: "Invoke /vibes:autoresearch". Do not decompose.
Prerequisites Check
VIBES_ROOT="${CLAUDE_PLUGIN_ROOT:-$(dirname "$(dirname "${CLAUDE_SKILL_DIR}")")}"
echo "Checking autoresearch prerequisites..."
test -f "$VIBES_ROOT/scripts/eval-ssr-check.ts" && echo "✓ Tier 1.5 SSR check" || echo "✗ Missing eval-ssr-check.ts"
test -f "$VIBES_ROOT/scripts/eval-harness.ts" && echo "✓ Tier 2 harness" || echo "✗ Missing eval-harness.ts"
test -f "$VIBES_ROOT/scripts/eval-parallel.ts" && echo "✓ Orchestrator" || echo "✗ Missing eval-parallel.ts"
test -f "$VIBES_ROOT/scripts/eval-scoring.ts" && echo "✓ Scoring" || echo "✗ Missing eval-scoring.ts"
test -f "$VIBES_ROOT/eval/config.md" && echo "✓ Config" || echo "✗ Missing eval/config.md"
test -f "$VIBES_ROOT/eval/napkin.md" && echo "✓ Napkin" || echo "✗ Missing eval/napkin.md"
ls "$VIBES_ROOT/eval/specs/"*.md 2>/dev/null | wc -l | xargs -I{} echo "✓ {} eval specs found"
cd "$VIBES_ROOT/scripts" && bun -e "import React from 'react'; console.log('✓ React available')" 2>/dev/null || echo "✗ React not installed"

If any prerequisite is missing, stop and inform the user.

Running
Option 1: Full Autonomous Run (Recommended)

Dispatch the autoresearch orchestrator agent (.claude/agents/autoresearch-orchestrator.md) with context from eval/config.md, eval/napkin.md, and eval/scoreboard.md.

Pass any CLI arguments from the user (e.g., --variants=5 --generations=10).

Option 2: Single Eval Pipeline Test
VIBES_ROOT="${CLAUDE_PLUGIN_ROOT:-$(dirname "$(dirname "${CLAUDE_SKILL_DIR}")")}"
bun "$VIBES_ROOT/scripts/eval-parallel.ts" --mode=eval-only <app.jsx> <spec.md>
Option 3: Score Existing Generation
VIBES_ROOT="${CLAUDE_PLUGIN_ROOT:-$(dirname "$(dirname "${CLAUDE_SKILL_DIR}")")}"
bun "$VIBES_ROOT/scripts/eval-scoring.ts" "$VIBES_ROOT/eval/results/gen-N/"
What It Does

Each generation:

  1. Mutate: N independent SKILL.md variants (fix-targeted, structural, adversarial, etc.)
  2. Generate: Each variant × each prompt × 3 runs
  3. Evaluate: Tier 1 (static) → Tier 1.5 (SSR) → Tier 2 (data model) — all programmatic
  4. Score: Triple-run averaging with consistency penalty; fitness = mean - 0.5×stddev
  5. Select: Best variant replaces current SKILL.md; git commit on improvement
  6. Repeat: Until plateau (3 gens), max generations, or score oscillation
Monitoring
  • eval/results/gen-N/summary.json — per-generation results
  • eval/results/summaries.json — cumulative history
  • eval/scoreboard.md — human-readable scoreboard
  • eval/napkin.md — failure log (grows monotonically)
Final Report
VIBES_ROOT="${CLAUDE_PLUGIN_ROOT:-$(dirname "$(dirname "${CLAUDE_SKILL_DIR}")")}"
bun "$VIBES_ROOT/scripts/eval-report.ts" "$VIBES_ROOT/eval/results/summaries.json"
按 MIT 许可原样转载,未经改动 · 在 GitHub 查看 →

评论

登录即可评论;带「已验证安装」的,是发布者名下有本店的安装或持有记录。
…