eval

Run Full Evaluation Pipeline

Run the complete benchmark pipeline: execute → judge → report.

echo "Arguments: $ARGUMENTS"

Pipeline Stages

1. Execute: For each test case × target, spins up a sandboxed VM, has an AI agent solve the problem, extracts solution files
2. Judge: For each test case × target, an LLM judge compares the generated solution to the reference solution and scores it
3. Report: Aggregates all judge scores into a terminal scorecard and writes report.json

Options

• --resume: Resume from the last checkpoint of an interrupted pipeline
• --fresh: Only useful with --resume. Resets pipeline state so the run re-executes from scratch in the same run directory. Does NOT delete result files. Without --resume, a new run always starts fresh anyway.
• --label <name>: Human-readable label for this run
• --run <runId>: Only used with --resume. Target a specific run instead of auto-detecting the latest incomplete one.

Detecting Pipeline Status

Before running, you can check if a pipeline is paused/interrupted by reading the pipeline state file:

Pipeline state location: <project>/results/<runId>/pipeline-state.json

{
  "stage": "execute",
  "startedAt": "2026-04-25T10:30:00.000Z",
  "testCases": 15,
  "completed": {
    "execute": { "node-20": ["TC-001", "TC-002"] },
    "judge": { "node-20": [] }
  }
}

How to check status:

• stage is "execute" or "judge" → pipeline is incomplete/paused
• stage is "report" → pipeline completed successfully
• Compare completed[stage][target].length vs testCases to see progress
• No report.json in the run directory → pipeline didn't finish
• List runs: look for subdirectories in results/ containing run.json

Run manifest (results/<runId>/run.json):

{
  "id": "run-2026-04-25T10-30-00-000Z",
  "createdAt": "2026-04-25T10:30:00.000Z",
  "targets": ["node-20"],
  "testCount": 15,
  "label": "baseline v2"
}

How Resume Works

When --resume is passed:

1. Finds the latest incomplete run (where stage !== "report"), or uses --run <id>
2. Loads the saved pipeline state
3. Skips completed stages entirely (e.g., if stage="judge", execute is skipped)
4. Within a stage: only runs tests not yet in the completed map for each target
5. Progress is saved after each individual test — safe against crashes

Abort Handling

• First Ctrl+C: Graceful — finishes current test, saves state, prints "use --resume to continue"
• Second Ctrl+C: Hard exit — immediate process termination

Running the Pipeline

Run agentic-usability eval -p $ARGUMENTS and monitor the output. If interrupted, suggest --resume to continue.

For detailed pipeline internals, see pipeline-guide.md.