Benchmarks beamscope against a real grep/read baseline, for one or more repos of your choice — self-serve, repeatable, no external Python/tiktoken dependency.
mix beamscope.benchmark --repo /path/to/repo
mix beamscope.benchmark --repo /path/to/repo1 --repo /path/to/repo2
mix beamscope.benchmark --repo /path/to/repo --output docs/benchmarks/For each --repo, auto-discovers representative get_callers/
get_callees/find_call_path/search_code tasks from the repo's own
call graph (see Beamscope.Benchmark.TaskDiscovery — no hardcoded
per-repo function names, works against any repo), measures real token
counts for both the baseline and beamscope (Beamscope.Benchmark.Tokenizer
— a vendored cl100k_base-equivalent tokenizer via the tokenizers Hex
package, cross-checked against real tiktoken — no Python venv), and
times both with Benchee for a real, repeatable latency comparison.
Writes one timestamped Markdown report per repo to --output (default
docs/benchmarks/) as benchmark_<repo_name>_<timestamp>.md — re-running
never silently overwrites a previous result, so you can track how numbers
change over time.