/deep-review — random-exploration × cross-agent

Self-review (inline, after each bead) catches most bugs in the code an agent just wrote. Deep review catches the boundary bugs: code one agent wrote being called wrong by another agent, integration mismatches, and latent bugs in code no-one's been staring at.

Prompts

Use the verbatim prompts at .flywheel/prompts/deep-review.md, alternating Round A and Round B.

How to run

• Pick 1–2 agents who just finished a bead. Don't stop the whole swarm.
• Send Round A to one, Round B to another.
• After they report, swap: send Round A to the second agent, Round B to the first.
• Continue until two consecutive rounds come back clean from different agents.

What the two rounds find

• Round A (random exploration): latent bugs in utility modules, error paths, and edge branches that nobody has reviewed because nobody's been touching them.
• Round B (cross-agent): boundary mismatches between agents' code, root causes behind symptoms, "I trusted the other agent's assumption and they weren't thinking about X."

If agents keep finding bugs after 4+ rounds

Back to bead space. Create fix-beads for the bug classes that keep appearing. The real problem isn't the individual bug, it's the pattern.