analyze-comparison-tests

name	analyze-comparison-tests
description	Collects comparison test run artifacts and answers the user's questions based on the trajectories of each run. WHEN TO USE: collect comparison test artifacts
license	MIT
metadata	{"author":"Microsoft","version":"1.0.0"}

Steps

1. Collect run artifacts

Execute the collect-artifacts script to download the test run artifacts.

The user must provide an JSON file to correlate each comparison test run with the GitHub Actions run. The script expects one input argument as the path to this JSON file. The JSON input is supposed to be the JSON output when queuing the comparison test runs using the npm run compare:run command.

cd tests/
npm run compare:collect -- input.json

The collect-artifacts script will download the test run artifacts to a directory named comparison-artifacts in the current working directory. Before executing the script, check if there is already such an directory. If so, skip executing the script and proceed to step 2.

2. Extract insights

The downloaded artifacts will have the following folder structure:

comparison-artifacts/
├── <branch-name>/
│   ├── <stimulus-name-1>/
│   │   ├── <model>-with-skill/
│   │   │   ├── agent-metadata-<date-string-1>.md
│   │   │   ├── agent-metadata-<date-string-2>.md
│   │   │   └── ...
│   │   └── <model>-without-skill/
│   │       ├── agent-metadata-<date-string-1>.md
│   │       ├── agent-metadata-<date-string-2>.md
│   │       └── ...
│   └── <stimulus-name-2>/
│       ├── <model>-with-skill/
│       │   └── agent-metadata-*.md
│       └── <model>-without-skill/
│           └── agent-metadata-*.md
└── <branch-name-2>/
    └── ...

Each <branch-name>/<stimulus-name>/<model>-with-skill or <branch-name>/<stimulus-name>/<model>-without-skill directory contains the test run trajectories for that stimulus and model on that branch, with or without skills. Each trajectory is a markdown file that records user prompts, tool call requests, tool execution results, assistant responses that happened during the run. It also contains statistics such as token usage and turns. Based on the trajectories, answer the user's questions for each test run. Generate a report following the report-template to show your answers.