hermes - 💡(How to fix) Fix Checkpoint backends: shadow-git vs CoW clonefile — has this tradeoff been considered?

I maintain a small project (cowback) that does roughly what tools/checkpoint_manager.py does — transparent per-turn snapshots of the working directory, rollback on demand — but backed by filesystem Copy-on-Write (APFS clonefile() on macOS, FICLONE ioctl on BTRFS/XFS) rather than shadow git repos.

Writing this up as an issue rather than a PR because I don't want to propose replacing something that works; I'm more curious whether a backend choice has been considered, and whether sharing numbers would be useful.

Quick comparison

Rough numbers (from my benchmark on APFS, M-series Mac)

(Shadow-git numbers are rough — I haven't run them against checkpoint_manager.py exactly; those are from running git add -A && git commit from scratch on synthetic trees. A real comparison would need to match the shadow-repo setup.)

Where each wins

Shadow git is clearly the right default for a cross-platform agent: works on every filesystem, brings dedup across long sessions, uses well-understood tooling. For a project Hermes's size this matters — you don't want to ship a backend that breaks on someone's ext4.

CoW wins when:

• The user has an APFS / BTRFS / XFS-reflink filesystem (most Mac devs, many Linux setups)
• Projects contain large binaries / assets / lockfiles that change frequently
• Snapshot latency per turn matters (LLM-speed loops benefit from ms-level snapshots)

Possible directions (if useful)

1. Nothing changes. The current impl is well-engineered and the perf is fine for most users. Totally reasonable answer.
2. Optional backend. A HERMES_CHECKPOINT_BACKEND=cow env / config flag that uses copyFileSync(..., COPYFILE_FICLONE_FORCE) when supported, falls back to shadow git otherwise. Extra ~100 lines of code.
3. Just the perf learnings. Even without a backend switch, some of the optimizations (e.g., skipping hash on large files) might be worth porting the other direction.

Happy to write up (2) as a PR against optional-skills/ or tools/ if there's interest — but only if the direction makes sense to the maintainers. If not, I'm happy to hear why; shadow-git has real advantages (hash dedup over long conversations) that I might be under-weighting.

Thanks for the well-written checkpoint_manager.py — reading it was the clearest "here's what agent checkpointing should look like" document I've seen.