Skip to main content

Entry Points

The 0.6.0 release publishes both CLI entry points. Run them from a v0.6.0 source checkout:
BenchMain registers Agent Judge’s deterministic file, build, and coverage judges. During run and resume, its test-quality-llm registration abstains because no model-backed judge is present. BenchApp replaces that registration with TestQualityJudge, backed by Agent Client and the Claude agent implementation. Other commands delegate to the core CLI. All benchmark setup, post-processing, and configured agent commands execute as local host processes with the invoking user’s permissions. Workspaces do not provide isolation.

Discovery Commands

list

tasks

--benchmark is required.

End-to-End Commands

run

Prepares each selected task, runs setup scripts and the configured agent, runs post-processing, and grades the result.

resume

Resumes runs/<uuid>, preserving completed trial results and running missing trials.
--run-id is required.

compare

--runs accepts one or more run-directory paths and prints accuracy, pass@k, cost, duration, and trial-count comparisons.

Split Workflow

provide

Creates the workspace, copies an optional template, and writes INSTRUCTION.md and .bench-context.yaml.

grade

Evaluates an existing workspace with the benchmark’s deterministic Agent Judge configuration.
Both commands require --benchmark and --workspace; --task is optional only for a single-task benchmark. The built-in split CLI is for deterministic juries. Use agents-module run or resume when the jury includes the real test-quality-llm judge.

Output Structure

Trial failures use the result model’s FailureMode values, including agent timeout/error, setup or grade errors, build/test failures, context exhaustion, and unknown failures.