Claude Code, Codex and Cursor crash recovery

First check what the crash left behind

From the folder containing .nightshift/, read the open boxes in punch-list.md and the latest shift-log.md entry. Ask Nightshift for Status and Doctor in the same project. If STOP, a completed shift, a deadline, or a clean close is recorded, the watchman should stand down; Start is the deliberate way to re-arm a paused contract.

# From the Nightshift workspace on macOS or Linux
ls .nightshift/STOP .nightshift/.shift-armed .nightshift/.ended 2>/dev/null
tail -40 .nightshift/shift-log.md

How Claude Code, Codex, and Cursor recover

A recorded identity and positive failure evidence matter more than a quiet screen. The watchman runs outside the host process and advances a lease before a replacement can use tools.

• Claude Code: transcript, process, roster, clean-exit, Escape, and API-error signals guide revival. It tries the recorded conversation first, then documented fallbacks; a clean close or Escape is not treated as a crash.
• Codex: pulse, process, rollout, and SessionEnd are decisive. SessionEnd pauses recovery; a crash without it may revive. Same-conversation resume needs a recorded resumable task identity. Known non-resumable handles stand down.
• Cursor: a dead IDE-owned shift can move to a new CLI worker with the same contract. The IDE conversation ID cannot resume in the separate CLI store. Later recoveries resume that CLI worker. Cursor CLI requires project hook installation when Setup offers it.

If the old panel still shows an error

Read shift-log.md for a resume attempt and inspect the punch list and receipts. A headless Claude Code or Codex worker can append turns without refreshing an already-open panel. Reopen the recorded conversation before prompting; the process lease fences old observable tool calls. Cursor's replacement CLI worker is a different conversation, so follow the recorded attach instruction instead of treating the IDE tab as current.

What recovery cannot do: sleep, power-off, stopped containers

A powered-off or sleeping machine cannot run the local watchman; a stopped container or remote environment also stops its local processes. A live but unclassified hang is left alone to avoid duplicate workers. An upstream outage may delay every retry. If Doctor reports missing or unsafe identity, keep the open contract and restart attended rather than manufacturing a resume.