The problem
One long agent session loses track of a big task: context fills up, mistakes carry forward, and afterwards it is hard to see why anything happened.
Next: What I didWhat I did
I split the work into roles, each running on its own model, and made the filesystem the shared memory:
The planner writes plan.md and one file per task, with dependencies, verify commands and acceptance criteria.
Tasks run in dependency order, each in a new session. On failure the error is fed back for a retry, then the task escalates, then it is marked failed.
A reviewer writes reviews/final.md ending in PASS or FAIL. Failures go back to the planner, up to a set number of rounds.
Because each handoff is a file and not a prompt, every step can be inspected after the run. The code is six source files.
Next: Where it standsWhere it stands
A working prototype, run from the command line.
It has no tests and no public repository yet.
Next: What I'd do nextWhat I'd do next
Add tests. The retry, escalation and replan paths are where a bug would cost the most.
Measure it. Comparing runs against a single session would show whether the roles earn their extra cost.