skill-forge-benchmark

Skill Benchmarking & Performance Tracking

Measure and compare skill performance across iterations with statistical rigor using multiple trials, variance analysis, and trend tracking.

Process

Step 1: Define Benchmark Configuration

Accept configuration as:

• Existing eval set: Path to evals/evals.json (from /skill-forge eval)
• Benchmark config: Custom config with trial count and thresholds

Benchmark config schema:

{
  "skill_name": "my-skill",
  "skill_path": "./my-skill",
  "eval_set_path": "./evals/evals.json",
  "trials_per_eval": 3,
  "baseline_type": "no_skill",
  "previous_benchmark": null,
  "thresholds": {
    "min_pass_rate": 0.8,
    "max_avg_tokens": 100000,
    "max_avg_duration_seconds": 120,
    "min_improvement_ratio": 1.0
  }
}

Step 2: Execute Benchmark Runs

For each eval, run trials_per_eval times (default: 3) to get reliable metrics:

1. Execute with-skill runs (3x per eval)
2. Execute baseline runs (3x per eval)
3. Capture per-run: pass/fail, token count, duration
4. Save each run's timing.json and grading.json

Use agents/skill-forge-executor.md for parallel execution where possible.

Step 3: Aggregate Results

Run python scripts/aggregate_benchmark.py <workspace>/iteration-<N> --skill-name <name>:

Output benchmark.json schema:

{
  "skill_name": "my-skill",
  "iteration": 1,
  "timestamp": "2026-03-06T12:00:00Z",
  "summary": {
    "total_evals": 10,
    "with_skill": {
      "pass_rate": 0.87,
      "pass_rate_std": 0.05,
      "avg_tokens": 45000,
      "avg_duration_seconds": 34.2
    },
    "baseline": {
      "pass_rate": 0.60,
      "pass_rate_std": 0.08,
      "avg_tokens": 62000,
      "avg_duration_seconds": 52.1
    },
    "improvement_ratio": 1.45,
    "token_savings_ratio": 0.73,
    "time_savings_ratio": 0.66
  },
  "per_eval": [
    {
      "eval_id": 0,
      "eval_name": "basic-trigger",
      "with_skill": {"pass_rate": 1.0, "avg_tokens": 30000, "avg_duration_seconds": 20.1},
      "baseline": {"pass_rate": 0.67, "avg_tokens": 50000, "avg_duration_seconds": 45.0},
      "trials": 3
    }
  ],
  "thresholds_met": {
    "min_pass_rate": true,
    "max_avg_tokens": true,
    "max_avg_duration_seconds": true,
    "min_improvement_ratio": true
  }
}