Ten Claude Code sessions, 1,138 merges, one solo dev's failure log

The setup was deliberately layered: a human decides what to build and approves production changes, an interface session relays work, an orchestrator sorts and verifies issues without implementing, and sessions A through G each live in their own git worktree and branch. A Windows machine handles GitHub Actions and deployment watching. The structure grew from just an orchestrator and two workers in early September into its present form as problems appeared.

The log of what went wrong is unusually concrete. A rule with no author appeared in a summary and authorized real merges. A documentation-only pull request skipped CI and turned main red because it contained four paths that did not exist. Five rapid merges on the night of 9/7 left four deployments cancelled with zero jobs, because GitHub Actions concurrency only holds one running and one waiting job, and cancel-in-progress: false did not help. A backend sat 27 commits behind production while showing green, because Docker Desktop was waiting on an admin password and a diff check compared only against the last push. Five sessions exhausted GitHub's account-wide rate limit and then read empty error responses as 'done'. git stash, shared across worktrees because there is only one .git, dropped another worker's changes. A stale worktree nearly reverted two features that had landed on main. Daily checks vanished because they lived in a scratchpad wiped on restart. Two sessions both named 'orchestrator' appeared in the list, one of them actually being worker F.

His own conclusion is that four causes explain all of it: unnoticed shared resources, important things kept in places that get erased, failing to distinguish absence from a failed fetch, and never confirming who decided what. What remains is a handful of habits rather than tooling — record decisions in commit messages, file an issue the moment something is requested, measure numbers on the spot, keep the orchestrator out of implementation, and halt production changes right before execution. He also noted that one X post on 9/27 brought 1,348 visitors, of whom three in four left without solving a single question. Whether that funnel fixes faster than more sessions add breakage is likely the real test of the approach.