[gouveags]

Related issue: #854.

Implement Smithy's requestCompression trait so generated clients compress eligible request bodies automatically. The public controls follow the shared AWS SDK configuration: disable_request_compression and request_min_compression_size_bytes (default 10240; inclusive range 0 through 10485760). Setting the threshold to zero compresses eligible buffered requests regardless of size; operations without gzip compression metadata remain uncompressed.

smithy-http owns the compression policy, smithy-core exposes the typed trait, and smithy-aws-core resolves standard configuration and integrates the policy with AWS protocols. Generated AWS clients inherit the controls; non-AWS generated HTTP clients receive the standard properties through the generator.

Compression runs after serialization and before signing. Wire headers reflect the compressed body, buffered-request retries reuse the same compressed bytes, and seekable streaming sources support hashing and rewinding without buffering the entire compressed payload. Event streams and streams requiring a known length remain excluded. Existing streaming-input retry restrictions remain.

The implementation also preserves buffered async-iterator data correctly on rewind, closes partially consumed gzip generators, and makes the aiohttp empty-body probe read one byte and restore position, including error and cancellation paths. Regression tests cover these transport boundaries.

Request compression can reduce upload bytes for conversation-heavy agents; it does not reduce model tokens or establish an end-to-end latency improvement. Bedrock Runtime still needs service acceptance and upstream compression metadata. Mantle's provider-compatible APIs need separate API/client coordination. No unsupported service operations are manually marked as compression-capable. Gzip is supported under the current contract; zstd requires separate contract and service agreement.

Related contributions: new SDK adoption and botocore threshold-range compatibility fix. Botocore already has its own compression implementation and does not depend on this PR.

Verification

Hosted CI on the final head reports action_required with no jobs run. Maintainer action is needed before hosted CI can validate this revision.

Reviewed implementation: 1ae921bac24a6ecebc61b8b43cd36382d359b4ec.

• The original HTTP compression regression failed before implementation. Existing regressions cover compact trait values, request signing, buffered retries, stream seeking and cancellation/cleanup. Removal of the custom configuration field was also checked before and after narrowing the API.
• make test-py: 2740 passed, 10 skipped.
• make test-protocols: 783 passed, 32 expected failures across seven generated projections. Request-compression protocol cases are enabled.
• make check-py: Ruff lint/format and strict Pyright passed.
• make build-py, make build-java and Java formatting checks passed.
• make generate-protocol-tests: passed after committing the regenerated configuration snapshot; generation reproduces the committed output.
• git diff --check passed.

Validation used Python 3.12.14, Java 17 and Pandoc 3.8.2. The final follow-up commit changes only release-note wording; runtime/generator evidence is unchanged from 800d40ecc45e2f134a38a9625ba24dd46725d3c6. Tests are hermetic; no live endpoint acceptance or latency validation is claimed. The full supported-Python matrix and hosted CI are evaluated separately. Platform E2E changes are not needed: runtime, real-signer/transport regressions and generated protocol tests exercise the changed contract.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.