A disposable measurement surface for the cache prepared by KnickKnackLabs/shimmer#797 and consumed by ricon-family/fold#159.
The workflow uses Fold's exact tool declaration and reviewed cache key with only the ephemeral repository token. Sequential workflow dispatches measure the first install/upload and subsequent restore paths without agent credentials or home preparation.
Measured on 2026-07-16 from signed commit
45fbe768e2fb0d65dbee4441c229b16714ba0b09:
| Path | Run | Cache step | Whole job | Result |
|---|---|---|---|---|
| Cold install and save | 29523585765 | 118s | 126s | cache_hit=false |
| Warm restore | 29523760569 | 11s | 19s | cache_hit=true |
The warm path saved 107 seconds, about 91% of the cache-step time. GitHub stored
652,617,103 bytes (about 622 MiB compressed) for 2,052,644 KiB and 24,296 files
on disk. Restore downloaded at about 109 MB/s and completed successfully. Both
runs passed the cache-boundary audit: no ephemeral repository token,
credentialed Git remote, or external symlink was present in MISE_DATA_DIR.
This proves the exact Ubuntu tool-cache mechanism and configuration used by the
candidate Fold workflow. It does not include later credential, home, or
ci:env preparation, so a post-merge agent wake remains the end-to-end latency
check.