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"