What changed
v0.2 is a pivot, driven by measured data.
In our own end-to-end run, a delegated task passed 16/16 tests after 2.5 minutes of worker time — and then the manager (Codex) spent another 5.5 minutes and 80% of the total tokens re-reading, re-editing, and re-testing the code.
The expensive part wasn't delegation. It was verification without a stopping rule.
What's new:
- Verification contract in the worker brief: allowed paths, checks, invariants, limits, review mode.
scripts/verify_workspace.sh— a non-LLM verifier that checks the workspace against the contract and prints a small JSON receipt.- Risk-tiered review:
receipt/targeted/full. - Hard stop rule: contract passes → report completion. No re-editing, no scope creep.
Core rule, extended:
Delegate the work. Verify the evidence. Don't redo the work.