bernstein
“The fix was to make the agent prompt explicitly instruct every agent to run BOTH ruff check AND ruff format --check before pushing, closing the gap where a partially-specified lint step let format violations reach main across the parallel agent fleet.”
solo-developer
Why they built it
The author wanted to run parallel multi-agent work directly on production codebases without the agents thrashing or coupling through shared memory, so they built a deterministic Python dispatcher where state lives in files and the orchestrator is code-driven rather than model-driven.
What worked
- ✓Strong test discipline held as the suite grew from 2,000 to 2,360 tests with zero breakage, module boundaries were enforced, scope discipline prevented silent version bumps and README rewrites, and per-agent git-worktree isolation kept ten concurrent agents from colliding.
What broke or was painful
- ✗When ten agents were dispatched against one release, six of the ten ran ruff check but never ran ruff format --check before pushing, so formatting drift slipped through the gate that everyone assumed was covering it; the author described it as obvious in retrospect and invisible until something forced it.
The result
The fix was to make the agent prompt explicitly instruct every agent to run BOTH ruff check AND ruff format --check before pushing, closing the gap where a partially-specified lint step let format violations reach main across the parallel agent fleet.
What they'd do differently
Specified the full lint contract (check plus format-check) in the shared agent prompt instead of assuming agents would infer the formatting step from a generic lint instruction.