The problem
Coding agents are quick, but they mark their own homework. An agent that wrote a change is the worst reviewer of it: it shares every assumption it made while writing it.
I was also doing the same steps by hand on every task. Agree what done looks like, build it, ask another model to review it, check the review was right, open a PR, watch CI. I wanted that written down once, so every agent on every machine follows it.
What I built
cstack is a set of skills that Claude Code and Codex load by name. One command installs them on a machine, and one repo serves every machine I use.
- GoalThe agent writes the goal, the problem and the smallest change it expects, then waits for my OK. After that the goal is frozen.
- Build or fix
cstack-buildlearns the code, builds, and proves it works in the app.cstack-fixreproduces the bug, proves the cause, commits a failing test, then fixes it. - Break loopTwo fresh adversaries, one Claude and one Codex, get the same brief and try to break the change. Every finding needs a repro.
- PR
full-stopfiles one PR with every round's findings, watches CI and review bots, and stops when it's ready to merge. Say "full send" and it merges once the checks pass.
How the break loop works
The agent that wrote the change is the developer. It never judges its own work alone. It reruns every repro the adversaries send. A finding that reproduces gets fixed, one commit each, or goes to an arbiter from the other model family. The developer can't drop a finding or call it out of scope by itself.
Then a new pair of adversaries, who know nothing about earlier rounds, gets the identical brief. The loop ends when a fresh round finds nothing real. There's no round limit.
- Fresh reviewers every round.A reviewer that saw round one starts trusting the fixes. A fresh one doesn't.
- No repro, no finding.Asking for proof cut out the review comments that sound clever but aren't true.
- Settle disputes across families.When Claude disagrees with a finding, Codex decides, and the other way round, so the two models' blind spots don't overlap.
What I borrowed
full-stop, full-send and parts of the break loop are adapted from Theo's tstack. A small file records the exact upstream commit for each one, and cstack upstream shows what changed there since I copied it.
What's next
- Log real runs, so the waterfall above comes from data: rounds per change, findings per round, and how many the arbiter upheld.
- Spot loops that stall, where each fix creates the next round's finding, and push the developer to a smaller fix sooner.
- Swap models from one config file as new ones come out. The current pair is Opus 5.5 and Sol 6.2.