CrashLabdocs
Reference

0.1 release status

What the local alpha supports and what it deliberately leaves out.

CrashLab 0.1 is an open-source alpha for one complete local development loop: set up an evaluation, run the real agent in Docker, understand the evidence, change the agent, and compare the candidate with a baseline.

Supported path

  1. Install CrashLab from crashlabsai/tap with Homebrew.
  2. Install the bundled coding-agent setup skill.
  3. Connect one repository-owned evaluation to the real agent.
  4. Run fresh trials in local Docker.
  5. Diagnose the retained result under .crashlabs/runs/.
  6. Run a baseline/candidate regression check.

The Homebrew formula installs the Python CLI, Ink workspace, uv, and their managed Python and Node runtimes on macOS Apple Silicon and Intel. It retains the exact release wheel so locked evaluation packages can create fresh runtimes without a separate PyPI release. The release archive also retains CrashLab's MIT license and the viewer's generated third-party notices. Every normal simulation cleans the ephemeral Docker resources it owns.

Capability map

Capability0.1 statusWhat that means
Local Docker executionSupportedDocker Engine/Desktop and Buildx are required.
run, focused reruns, and checkSupported alphaThey execute the configured real target and retain results automatically.
Interactive workspaceSupportedBare crashlab opens coverage, runs, and past evaluations without a second execution engine.
Python/uv, Python/requirements, and Node/pnpm source shapesSupported alphaSetup still has to connect a real repository agent, world, and verifier.
Deterministic protected verificationSupported alphaWorld state, invariants, and causal events stay outside the target.
Semantic verificationSupported alphaA customer-selected protected judge can evaluate bounded evidence; uncertainty never silently passes.
Reliability, latency, tokens, and costSupported alphaMissing provider data is shown as unavailable, not zero.
Remote sandboxesNot includedDocker is the only supported 0.1 execution backend.
Linux and Windows packagesNot includedThe initial Homebrew package is macOS-only.

What setup proves

A generated folder is not a completed setup. The evaluation must import an existing source-controlled agent, run it against a functional stateful world, and retain at least one protected result. CrashLab refuses to treat a scripted control or an agent invented inside the evaluator as customer evidence.

Clean distribution exercises have completed that loop against unrelated Python and TypeScript/JavaScript applications. They also caught a semantic-only target regression, guided a coding agent to the failed behavior, and passed a repaired baseline/candidate comparison. These exercises support the local workflow; they do not make one trial statistically conclusive for every agent.

Alpha boundaries

CrashLab 0.1 does not claim that:

  • crashlab init can infer an arbitrary product workflow by itself;
  • a source scaffold is a finished evaluation;
  • one passing trial proves production reliability;
  • unavailable required verification can be skipped;
  • plausible final text proves that a tool or mutation executed;
  • unreviewed Python evaluation packages are safe to import; or
  • current alpha schemas are stable 1.0 contracts.

Evaluation packages should import from crashlab.alpha, keep a current uv.lock, and run crashlab eval validate during upgrades. Use crashlab view <stored-run> --json for a bounded diagnosis and --raw only when you need the complete authoritative evidence.

Report a release-path problem

Include the CrashLab version, host platform, exact command and exit code, retained selector or redacted diagnosis, and Docker/Buildx versions. Remove credentials, customer source, and sensitive evidence before opening an issue in the CrashLab repository.

On this page