Benchmark launcher

Launcher-side integration for benchmark runs. The workspace member publishes bro-bench, and its package is bro.bench, a portion of the framework's bro namespace. It carries only code the root environment must install so launches can register benchmark credentials, the benchmark worker type, and session commands without importing Harbor.

Development

This member depends on core and joins the root workspace environment. Formatting, linting, typing, tests, and packaging use the root repository gate. Run sync-scripts --project bench after adding or removing a CLI, and build the wheel with uv build --package bro-bench.

Components

• bro/bench/credentials.py — the harbor and benchmark_retention credential-kind declarations.
• bro/bench/job.py (benchmark-job) — the registered BenchmarkType, its timeout option declaration, and the session client that starts or checks a raw Harbor run through launch. A launch is denied when its config's environment is Harbor's docker one and the host lacks the compose plugin that environment drives.
• bro/bench/presets.py — the presets under benchmark/ (agents, settings, datasets, tasks): their shapes, and their composition into a Harbor job config carrying the names it came from.
• bro/bench/run.py (benchmark-run) — the operator loop around one run: it composes the presets, builds the trial bundle, runs the job through the session broker or, outside a session, as a host process into a jobs directory, reports it, and retains it on request.