Retry After an API Error / Outage

A transient Anthropic API failure (overload, 5xx, rate limit, network blip, or a full outage) aborts the turn mid-task. This skill cleanly resumes the work that got cut off — it does NOT start fresh and does NOT re-litigate decisions already made before the error.

First: is this actually a retry?

Only proceed if the last failure was infrastructure, not the work itself:

✅ Retry: Overloaded, 429/529, 500/503, rate_limit, fetch failed, socket hang up, “Claude is down”, the turn just stopped with no result.
❌ Not a retry — that’s a fix: a test failed, a build broke, wrong output, a logic bug. Don’t blindly redo it; diagnose. Say so and switch to fixing.

If it’s ambiguous, ask the user which it was before acting.

Procedure

Reconstruct where it died. Read the last few turns + any handoff/state (.omc/state/, .omc/handoffs/, todos, open files) to find the exact action that was in flight when the error hit. Don’t assume — verify against the actual repo/file state, since a partial write may have landed.
Check for partial work. git status --short and inspect the target file(s). The failed action may have half-completed (a file written but not verified, a commit started). Reconcile before redoing, so you don’t double-apply.
Re-run only the cut-off action, not the whole task. Pick up from the last verified-complete step. If several steps remain, restate the short remaining list first so the user can confirm scope.
Verify end-to-end before claiming done (build/test as the project requires) — the interruption means nothing downstream was checked.

Surviving a LIVE outage (the wrapper)

This skill can’t help while the API is actually down — the model isn’t running to invoke it. For that, use the shell wrapper, which relaunches the CLI with exponential backoff and resumes via --continue:

~/.claude/scripts/claude-retry.sh [any normal claude args]
# e.g.  ~/.claude/scripts/claude-retry.sh -p "finish the migration"
#       ~/.claude/scripts/claude-retry.sh           # interactive TUI

It retries only transient failures (overload/5xx/rate-limit/network), backs off 4s→8s→16s… (capped), and stops immediately on auth/credit errors. Suggest a shell alias for convenience:

alias claude-r='~/.claude/scripts/claude-retry.sh'

Tune via env: CLAUDE_RETRY_MAX (default 5), CLAUDE_RETRY_BASE (4s), CLAUDE_RETRY_CAP (120s). Log at ~/.claude/claude-retry.log.