# Issue **Title:** [v2] Update standard retry mode via opt-in env var ## Description *Notes:* * Ports to vendored botocore. The main and only difference from upstream botocore is that AWS CLI v2 uses `standard` retry mode by default, but upstream botocore uses `legacy`. *Description of changes:* * Introduced new `AWS_NEW_RETRIES_2026` environment variable. When set to true, standard retry mode behavior is changed. * The standard retry mode behavior changes (gated behind the new env var) is described below: * **Service-specific max attempts** — Service-defined (e.g., DynamoDB with 4 attempts) max attempt values instead of the universal default of 3, but only if the user hasn't explicitly configured max attempts. * **Throttling-aware backoff** — The ExponentialBackoff is initialized with a `throttling_detector` and `service_name`, enabling it to use different backoff scales depending on whether the error is a throttling error and which service is being called (e.g., 25ms base for DynamoDB non-throttling errors vs. 50ms for other services vs. 1s for throttling errors). * **Throttling-aware retry quota** — The RetryQuotaChecker receives a throttling_detector, allowing it to deduct a smaller quota cost (5) for throttling errors vs. the standard cost (14) for other retryable errors. * **Service name passed to RetryHandler** — For known long-polling operations (e.g., `SQS ReceiveMessage`, `SFN GetActivityTask`), when retry quota is exhausted the handler sleeps the backoff delay before returning without retrying. * Added new test cases for the new standard retry mode behavior updates. *Description of tests:* * Successfully ran pre-production build workflow (see Build Internal GitHub Action). By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice. ## Task Modify the repository so that the issue described above is resolved. The repository is checked out at base commit `bdf748564bd6`. Edit files in place; the verifier captures your changes via `git diff` and scores them against an oracle patch using SWE-RL-style diff-similarity reward.