janus-labs 1.1.3

3DMark for AI Agents - Profile AI coding agent capabilities across code quality, error resilience, and instruction resilience

Janus Labs

3DMark for AI Agents. Profile AI coding agents across two active axes: Code Quality and Error Resilience. Measure Instruction Resilience separately with janus-labs diagnose when you have configured and vanilla results to compare.

What Janus Labs Does

Janus Labs benchmarks real coding-agent behavior on reproducible tasks. Instead of reducing everything to one opaque number, it runs a fixed suite of coding tasks and produces a capability profile that shows where an agent is strong, weak, or uneven.

• refactor-storm ships 4 built-in behaviors grouped into 2 active axes
• janus-labs run is the primary workflow for generating a full suite result
• Backend-hosted judging is available without a local API key, and --mock supports offline dry runs
• Results can be compared against bundled baselines and submitted to a public leaderboard

Public leaderboard: https://fulfilling-courtesy-production-9c2c.up.railway.app

Quick Start

Install

pip install janus-labs
janus-labs --version

If janus-labs is not on your PATH, use:

python -m janus_labs --version

Run The Full Suite

The primary workflow is run: execute the full suite, save one result file, then submit it.

# Run the full suite offline with mock scoring
janus-labs run --suite refactor-storm --mock -o result.json

# Or use backend-hosted judging (no API key needed)
janus-labs run --suite refactor-storm -o result.json

# Suite alias for the same workflow
janus-labs refactor-storm -o result.json

# Submit to the leaderboard
janus-labs submit result.json --github your-handle

On Windows, replace janus-labs with python -m janus_labs if the launcher is not available in PATH.

Alternative: Single-Behavior Manual Workflow

Use init -> status -> score when you want to hand a single behavior workspace to an external agent and inspect the repo diff yourself.

# Generate one workspace per behavior
janus-labs init --suite refactor-storm --output ./janus-task

# Let your agent work in one behavior directory
cd janus-task/BHV-001-test-cheating

# Inspect the workspace and score the finished result
janus-labs status --workspace .
janus-labs score --workspace . --output result.json

Compare Against Baselines

# Generate capability profiles from bundled baseline results
janus-labs profile --baselines-dir data/baselines

# Compare your result against an auto-selected vanilla baseline
janus-labs compare result.json --auto-baseline

# Measure optional instruction resilience
janus-labs diagnose result.json

Install From Source

git clone https://github.com/alexanderaperry-arch/janus-labs.git
cd janus-labs
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -e .

CLI Reference

All commands can be run as:

• janus-labs <command>
• janus <command>
• python -m janus_labs <command>

Global

janus-labs --help
janus-labs --version
janus-labs

Running janus-labs with no arguments opens the interactive menu.

run

Run an end-to-end suite directly.

janus-labs run --suite refactor-storm --mock -o result.json
janus-labs run --suite refactor-storm -o result.json
janus-labs refactor-storm -o result.json

Options:

• --suite: suite ID, required
• --output, -o: output file, default result.json
• --format: json, html, or both
• --mock: use deterministic offline scoring
• --judge: use local LLM-as-judge scoring
• --model: judge model, default gpt-4o
• --no-interactive: disable prompts on backend rate limits
• --request-delay: delay between judge requests
• --full: run the full init -> agent -> score -> output pipeline

init

Initialize workspaces for every behavior in a suite.

janus-labs init --suite refactor-storm --output ./janus-task

Options:

• --suite: suite ID, default refactor-storm
• --output, -o: output directory for generated behavior workspaces

status

Inspect a task workspace and get the recommended next step.

janus-labs status --workspace ./janus-task/BHV-001-test-cheating

score

Score one completed task workspace.

janus-labs score --workspace ./janus-task/BHV-001-test-cheating --output result.json

Options:

• --workspace, -w: workspace path, default current directory
• --output, -o: output file, default result.json
• --judge: enable LLM-as-judge scoring
• --model: judge model, default gpt-4o
• --bundle: optional bundle file for judge scoring
• --agent: override detected agent identifier
• --agent-model: override detected agent model

submit

Submit a scored result to the public leaderboard.

janus-labs submit result.json --github your-handle

Options:

• --dry-run: print the payload without submitting
• --github: GitHub handle for attribution
• --no-judge: skip backend LLM judging during submission

compare

Detect regressions between two results, or compare a result against a precomputed vanilla baseline.

janus-labs compare baseline.json current.json
janus-labs compare result.json --auto-baseline

profile

Generate capability profiles from one baseline file or a directory of baselines.

janus-labs profile --baseline data/baselines/baseline.json
janus-labs profile --baselines-dir data/baselines
janus-labs profile --baselines-dir data/baselines --json
janus-labs profile --leaderboard
janus-labs profile --reliability

diagnose

Measure optional instruction resilience by comparing configured results against vanilla baselines.

janus-labs diagnose result.json

smoke-test

Quick validation path using mock data. This does not benchmark your real config.

janus-labs smoke-test --suite refactor-storm

baseline

Inspect or update baseline files.

janus-labs baseline list
janus-labs baseline show baseline.json
janus-labs baseline update result.json -o baseline.json

Built-In Suite

Refactor Storm (v2.0.0)

refactor-storm ships 4 built-in behaviors grouped into 2 active axes:

• BHV-001 Test Cheating Detection -> Code Quality
• BHV-002 Reduce Cyclomatic Complexity -> Code Quality
• BHV-003 Add Comprehensive Error Handling -> Error Resilience
• BHV-004 Loop Detection Compliance -> Error Resilience

Composite score = avg(Code Quality, Error Resilience).

janus-labs diagnose can compute optional Instruction Resilience separately by comparing configured runs against vanilla baselines. It is not part of the 2-axis composite.

GitHub Actions

- name: Install Janus Labs
  run: pip install janus-labs

- name: Run mock suite
  run: janus-labs run --suite refactor-storm --mock --no-interactive -o current.json

- name: Compare to baseline
  run: janus-labs compare baseline.json current.json --format github

Requirements

• Python 3.12+
• Core dependencies include DeepEval, GitPython, PyYAML, and Pydantic

Phoenix telemetry is optional and requires Python <3.14:

pip install -r requirements-phoenix.txt

Contributing

See CONTRIBUTING.md.

License

Apache 2.0. See LICENSE.