eval-recipes-runner

Purpose

Run Microsoft's eval-recipes benchmarks to validate amplihack improvements against baseline agents.

When to Use

• User asks to "test with eval-recipes"
• User says "run the evals" or "benchmark this change"
• User wants to validate improvements against codex/claude_code
• Testing a PR branch to prove it improves scores

Capabilities

I can run eval-recipes benchmarks to:

1. Test specific amplihack branches
2. Compare against baseline agents (codex, claude_code)
3. Run specific tasks (linkedin_drafting, email_drafting, etc.)
4. Compare before/after scores for PRs
5. Generate reports with score improvements

How It Works

Setup (One-Time)

# Clone eval-recipes from Microsoft
git clone https://github.com/microsoft/eval-recipes.git ~/eval-recipes
cd ~/eval-recipes

# Copy our agent configs
cp -r $(pwd)/.claude/agents/eval-recipes/* data/agents/

# Install dependencies
uv sync

Running Benchmarks

Test a specific branch:

# Update install.dockerfile to use specific branch
# Then run benchmark
cd ~/eval-recipes
uv run eval_recipes/main.py --agent amplihack --task linkedin_drafting --trials 3

Compare before/after:

# Test baseline (main)
uv run eval_recipes/main.py --agent amplihack --task linkedin_drafting

# Test PR branch (edit install.dockerfile to checkout PR branch)
uv run eval_recipes/main.py --agent amplihack_pr1443 --task linkedin_drafting

# Compare scores

Available Tasks

Common tasks from eval-recipes:

• linkedin_drafting - Create tool for LinkedIn posts (scored 6.5/100 before PR #1443)
• email_drafting - Create CLI tool for emails (scored 26/100 before)
• arxiv_paper_summarizer - Research tool
• github_docs_extractor - Documentation tool
• Many more in ~/eval-recipes/data/tasks/

Typical Workflow

When user says "test this change with eval-recipes":

1. Identify the branch/PR to test
2. Update agent config to use that branch:

   # In .claude/agents/eval-recipes/amplihack/install.dockerfile
   RUN git clone https://github.com/rysweet/...git /tmp/amplihack && \
       cd /tmp/amplihack && \
       git checkout BRANCH_NAME && \
       pip install -e .
   

3. Copy to eval-recipes:

   cp -r .claude/agents/eval-recipes/* ~/eval-recipes/data/agents/
   

4. Run benchmark:

   cd ~/eval-recipes
   uv run eval_recipes/main.py --agent amplihack --task TASK_NAME --trials 3
   

5. Report scores and compare with baseline

Expected Scores

Baseline (main branch):

• Overall: 40.6/100
• LinkedIn: 6.5/100
• Email: 26/100

With PR #1443 (task classification):

• Expected: 55-60/100 (+15-20 points)
• LinkedIn: 30-40/100 (creates actual tool)
• Email: 45/100 (consistent execution)

Example Usage

User says: "Test PR #1443 with eval-recipes on the LinkedIn task"

I do:

1. Update install.dockerfile to checkout feat/issue-1435-task-classification
2. Copy to eval-recipes: cp -r .claude/agents/eval-recipes/* ~/eval-recipes/data/agents/
3. Run: cd ~/eval-recipes && uv run eval_recipes/main.py --agent amplihack --task linkedin_drafting --trials 3
4. Report results: "Score: 35.2/100 (up from 6.5 baseline)"

Prerequisites

• eval-recipes cloned to ~/eval-recipes
• API key in environment: export ANTHROPIC_API_KEY=sk-ant-...
• Docker installed (for containerized runs)
• uv installed: curl -LsSf https://astral.sh/uv/install.sh | sh

Notes

• Benchmarks take 2-15 minutes per task depending on complexity
• Multiple trials (3-5) give more reliable averages
• Docker builds can be cached for speed
• Results saved to .benchmark_results/ in eval-recipes repo

Automation

For fully autonomous testing:

# Test suite for a PR
tasks="linkedin_drafting email_drafting arxiv_paper_summarizer"
for task in $tasks; do
  uv run eval_recipes/main.py --agent amplihack --task $task --trials 3
done

# Compare results
cat .benchmark_results/*/amplihack/*/score.txt