Claude Code cleans up crashed runs' uncommitted files

The write-up is drawn from the records of a small setup where Claude Code runs as a CEO role every four hours: Windows Task Scheduler wakes a Python launcher, the launcher starts claude -p, and the session picks work, delegates to subagents, and writes an end-of-run report. The keep-it-working premise is that each task is committed as it finishes, and that the launcher checks a lock file before starting so two runs never share one working tree.

What makes the recovery step hard is that leftover files are not all the same. A charter entry tells the session that files the launcher itself writes — instruction sheets, execution logs, failure and skip records, break notices — are not crash artifacts and should be committed as-is; a dated example shows an instruction sheet left untracked and committed with that reasoning. Human edits get a third lane: off-limits, logged, and handed back to a person.

The author is candid about limits. Deciding whether a prior run's work is complete is a judgment call, so countable finish criteria make it more mechanical. And because deletion is not permitted, untracked broken files simply remain until a human decides. The state file adds a second check: the board and the git history can disagree, so all three — tree, history, state — get compared.