unloop-bench
Unloop Bench
Optional local measurements alongside unloop and Superpowers. The core works unchanged without this add-on. Version 0.1.0; Python 3.9+ on macOS/Linux.
Read usage.md for the requested operation. Resolve scripts relative to this loaded skill directory. The command-line help is the flag reference; event-schema.md covers manual receipts and evidence tiers.
| User intent | Operation |
|---|---|
| “Benchmark this feature” | Bind explicit enrollment to its existing core store and feature; save the returned ID in the handoff |
| “Resume benchmarking this feature” | Inspect that ID; resume paused capture or explicitly reopen finished capture |
| “Pause benchmarking” | Persist pause immediately; keep prior history |
| “Report the benchmark” | Export the existing local snapshot; summarize counts and unknowns |
Installation is not enrollment. A fresh session needs the enrollment handoff or an explicit resume request; this release has no invisible lookup or runtime hooks. A feature's replacement PRs use the same enrollment, never a new sample.
During enrolled work, record concise source-linked actions at meaningful implementation/repair boundaries. Wrap only verification commands already authorized for the task; import completed core evidence at useful boundaries. Carry model/runtime/version when an actual source supplies them. Missing metadata stays unknown.
The wrapper runs argv once without shell evaluation. By default it retains metadata and passes output through. Full argv/output retention requires an additional explicit privacy decision using --retain-raw. Failed collection cannot justify retrying the command.