Speeding up the AI agent development loop - local snapshot approach

[nguyenyou]

Speeding up the AI agent development loop - local snapshot approach

Sharing an approach I've been using to speed up the code → compile feedback loop when working with AI coding agents (Claude Code, etc.) on Mill projects.

The out/ folder occasionally gets invalidated for various reasons, and rebuilding from scratch takes a while. I wanted a fast way to restore to a known-good state.

Why not remote cache?

• out/ can be multiple GBs
• Upload/download often slower than just rebuilding locally on a fast machine
• Mill doesn't yet have layout-agnostic cache keys for reliable portability

My approach: local APFS snapshots

On macOS, APFS cp -c creates instant copy-on-write clones with near-zero storage overhead.

~/.mill-out-cache/
  └── <project>/
      └── <commit-hash>/   # APFS clone of out/

• mill-cache snapshot — syncs to upstream/main, compiles, snapshots out/
• mill-cache restore — finds nearest ancestor commit, restores instantly (~60 secs)

I only snapshot clean main branch commits as known-good baselines. When things go sideways, instant restore and back to work.

Anyone else optimizing their agent workflow? Curious what approaches others are using to keep the iteration loop fast when working with AI agents on Mill projects.

[lefou]

Regarding the technical challenges (not the use of AI itself):

2. Remote caching - the title is misleading and it's a perfect use case to use it locally, e.g. set it up on the same machine. This should speed up caching, esp. when switching between branches often.

3. We already have an open issue about making the Mill cache relocatable. Make out/ folder contents (more) reproducible and filesystem layout agnostic #3660. I have two PRs open that solve most of the identified issues:

• Dynamically map known root paths in cache files #6031 - Maps all absolute paths to some well known roots like the project workspace or the coursier cache
• New mill.api.opt API for configuration data containing paths #6129 - Provided an new API to be use for configuration data that may contain paths such as compiler options, classpaths, env variables and so on.

Boths PRs don't bring much in isolation and the second currently breaks binary compatibility, but they should fix most of the open issues when applied together. AFAIK, there is still a need to change the Mill code generator for build files, as it currently emits absolute paths in the generated files.