Claude Code API Error 500: diagnose and recover

Confirm the 500 came from the service

Read the actual Claude Code error and service status before changing project files. In the Nightshift workspace, inspect Status or Doctor, open boxes in punch-list.md, and the latest shift-log.md entries. A clean close, owner Escape, STOP, deadline, or completed contract is a deliberate ending, not a failed recovery.

How Nightshift resumes Claude Code after a 500

The watchman runs outside Claude Code, normally waking every ten minutes unless configured otherwise. When process, transcript, and roster signals show a dead session or a classified API-error wedge, it tries the recorded conversation first, then documented continuation or fresh-session fallbacks. It rechecks liveness before further attempts and records each attempt. The outage itself still costs time, and missing or uncertain evidence makes it stand down.

• The on-disk list, decisions, and receipts survive even if conversation resume fails.
• A replacement inherits the original permission scope; it cannot widen it.
• The machine and its watchman must remain running.

Verify recovery without duplicating work

If the old panel still shows the error, check shift-log.md and the files before sending another prompt. A headless revival may not refresh that panel. Reopen the recorded thread to interact with it, then inspect the item receipts, diffs, checks, and open boxes. If Doctor says the session is paused or cannot be classified, resume attended from the durable contract.

tail -40 .nightshift/shift-log.md