Field notes

Challenges

We pick work from the desk, not from a backlog of imaginary users. The numbering matches the issues.

1. 03Test an agent without mocking the world it acts onStill open · Issue 03 · RooklightA harness that waits on real jobs, asserts on tool trajectories, and refuses to stub the filesystem. Caching fixtures and visual assertions are unsolved.
2. 02A library you can save into without thinkingStill open · Issue 02 · Save4laterAndroid capture that survives Samsung, a Room library that must never be wiped, and a backend graph that disagrees with the phone more often than it should.
3. 01A planner that thinks like a localGraduated · Issue 01 · Travel RuterScheduling against stale maps, routing, and hotel APIs. The tool exists. The language for “this time might be wrong” does not.