model-evaluation-benchmark（模型评估基准）

Model Evaluation Benchmark Skill

Purpose: Automated reproduction of comprehensive model evaluation benchmarks following the Benchmark Suite V3 reference implementation.

Auto-activates when: User requests model benchmarking, comparison evaluation, or performance testing between AI models in agentic workflows.

Skill Description

This skill orchestrates end-to-end model evaluation benchmarks that measure:

• Efficiency: Duration, turns, cost, tool calls
• Quality: Code quality scores via reviewer agents
• Workflow Adherence: Subagent calls, skills used, workflow step compliance
• Artifacts: GitHub issues, PRs, documentation generated

The skill automates the entire benchmark workflow from execution through cleanup, following the v3 reference implementation.

When to Use

✅ Use when:

• Comparing AI models (Opus vs Sonnet, etc.)
• Measuring workflow adherence
• Generating comprehensive benchmark reports
• Need reproducible benchmarking

❌ Don't use when:

• Simple code reviews (use reviewer)
• Performance profiling (use optimizer)
• Architecture decisions (use architect)

Execution Instructions

When this skill is invoked, follow these steps:

Phase 1: Setup

1. Read tests/benchmarks/benchmark_suite_v3/BENCHMARK_TASKS.md
2. Identify models to benchmark (default: Opus 4.5, Sonnet 4.5)
3. Create TodoWrite list with all phases

Phase 2: Execute Benchmarks

For each task × model:

cd tests/benchmarks/benchmark_suite_v3
python run_benchmarks.py --model {opus|sonnet} --tasks 1,2,3,4

Phase 3: Analyze Results

1. Read all result files: ~/.amplihack/.claude/runtime/benchmarks/suite_v3/*/result.json
2. Launch parallel Task tool calls with subagent_type="reviewer" to:

- Analyze trace logs for tool/agent/skill usage - Score code quality (1-5 scale)

1. Synthesize findings

Phase 4: Generate Report

1. Create markdown report following BENCHMARK_REPORT_V3.md structure
2. Create GitHub issue with report
3. Archive artifacts to GitHub release
4. Update issue with release link

Phase 5: Cleanup (MANDATORY)

1. Close all benchmark PRs: gh pr close {numbers}
2. Close all benchmark issues: gh issue close {numbers}
3. Remove worktrees: git worktree remove worktrees/bench-*
4. Verify cleanup complete

See tests/benchmarks/benchmark_suite_v3/CLEANUP_PROCESS.md for detailed cleanup instructions.

Example Usage

User: "Run model evaluation benchmark"Assistant: I'll run the complete benchmark suite following the v3 reference implementation.

[Executes phases 1-5 above]

Final Report: See GitHub Issue #XXXX
Artifacts: https://github.com/.../releases/tag/benchmark-suite-v3-artifacts

References

• Reference Report: tests/benchmarks/benchmark_suite_v3/BENCHMARK_REPORT_V3.md
• Task Definitions: tests/benchmarks/benchmark_suite_v3/BENCHMARK_TASKS.md
• Cleanup Guide: tests/benchmarks/benchmark_suite_v3/CLEANUP_PROCESS.md
• Runner Script: tests/benchmarks/benchmark_suite_v3/run_benchmarks.py

Last Updated: 2025-11-26 Reference Implementation: Benchmark Suite V3 GitHub Issue Example: #1698