Summary

A single 429 or 5xx from the provider currently aborts the whole Claude Code turn, because client.py raises on the first non-200 response. This adds a bounded retry with exponential backoff in ProviderClient so a throttled free-tier provider recovers instead of killing the session.

Why this matters

The README points people at free tiers like Groq and NVIDIA NIM, and those tiers rate-limit hard. CONTRIBUTING.md lists "rate limit headers" as a wanted contribution, which is the gap this closes. LiteLLM and claude-code-router both retry transient failures already, so this brings Backdoor to parity, and it does so without adding a dependency.

Changes

Both complete and stream now retry on 429 and 5xx responses and on transport errors, up to PROVIDER_MAX_RETRIES (default 3) with exponential backoff, honoring a Retry-After header when the provider sends one. Errors that will not fix themselves on a retry, like a 404 for a bad model name, still raise immediately rather than sleeping through three attempts. Streaming is the subtle case: it only retries before the first chunk is yielded, so a mid-stream failure surfaces to the caller and the SSE stream is never restarted with a duplicate message_start. Setting PROVIDER_MAX_RETRIES=0 restores the old fail-fast behavior.

Testing

Added tests/test_retry.py with five cases on the existing pytest-asyncio setup: a 429 followed by a 200 retries and succeeds, a persistent 503 gives up after the retry budget and surfaces the error, a 404 raises with no retry, a zero retry budget makes a single attempt, and Retry-After parsing covers the delta-seconds and fallback paths. All pass, with no new dependencies beyond asyncio and stdlib email.utils.