We now have a vision

[gbelinsky]

I've been running this project as an itch to scratch. I wanted an agent I could actually trust with ongoing work, one that wouldn't lose context, forget what it was doing, or require me to babysit every run. The agents I
tried either did too little or hid too much.

So I built PersonalAgentKit around a few things I kept wanting: local state I could inspect, runs that resume when they get interrupted, and a system that gets more useful the longer it runs rather than starting fresh
every time.

I finally wrote that down properly. It's in VISION.md at the root of the repo.

Short version: this is for one technically capable person who wants a system that improves through use and continues through failures. Not a team platform, not a hosted product, not autonomy theater.

What also shipped:

• Checkpoint and resume. Killed runs can be picked up where they left off.
• Reflection contract. Certain goal types require a written reflection before they're considered done.
• Orphan reconciliation. Stale running runs are detected and finalized automatically.
• Reader-facing run contract. Every run directory now has a README.md as the entry point, and every queued goal gets a .run file showing where it went.

The last one came directly from the vision work. The agent audited its own filesystem experience and shipped the fix.