Problem

When a model provider returns a 429 rate limit error, the current behavior is to fail the request and potentially fall back to another profile/model. There's no automatic retry with delay.

Current Behavior

• Rate limits are detected (FailoverReason: 'rate_limit')
• Profile gets marked with a cooldown (calculateAuthProfileCooldownMs)
• Request fails immediately

Proposed Behavior

For 429 responses:

1. Check Retry-After header if present
2. Apply exponential backoff: 1s, 2s, 4s, 8s, 16s (max 3-5 retries)
3. Retry the same request
4. Only fail/fallback after exhausting retries

Why This Matters

Rate limits are often transient. Batch operations that hit rate limits shouldn't fail entirely - they should wait and retry. This is especially important for:

• Long-running batch jobs
• Multi-model operations
• High-throughput scenarios

Implementation Notes

The retry logic should integrate at the API call level, possibly in:

• src/agents/pi-embedded-runner/run/attempt.ts
• Or a wrapper around the provider SDK calls

The existing calculateAuthProfileCooldownMs shows the backoff pattern already in use for profile cooldowns.

Prior Art

• OpenAI SDK has built-in retry
• Anthropic SDK has configurable retry
• Most HTTP clients support retry middleware