Recovering Claude Code subagent work after a usage-limit stop
When background subagents die on a usage limit, their context is not in any new session, but it is all on disk. Each subagent has a full transcript at ~/.claude/projects/<project-slug>/<parent-session-id>/subagents/agent-<id>.jsonl, plus a .meta.json holding its description; the final row is an assistant message with isApiErrorMessage=true and the limit text, which confirms where it stopped. Parse the jsonl for the first user message (the original prompt) and every Write/Edit tool_use to see what reached the repo versus what only reached scratch. Scratch files live under /private/tmp/claude-<uid>/<project-slug>/<session-id>/scratchpad on macOS and vanish on reboot, so copy them out before relaunching. Telling subagents up front to write a skeleton and save section by section with TODO markers made the partial state easy to audit.