agentjido / req_llm / 20394c1ace4b7b4de287faf16e42996b014812c8

49%

Build Type

push

github

Committed by web-flow

Commit Message

Fix Bedrock streaming, tool round-trip, and usage metadata for 1.0 (#147)

* Fix Bedrock streaming temperature/top_p conflicts and increase timeout

- Apply translate_options in attach_stream to handle parameter conflicts
- Streaming requests were bypassing Options.process pipeline
- Increase receive_timeout from 30s to 60s (measured at 48s for fixtures)
- Fixes HTTP 400 errors when streaming with both temperature and top_p

* Fix Jason.EncodeError when saving streaming fixtures with binary protocols

The "decoded" field in streaming fixtures was added for human readability
(raw text alongside base64). However, some providers use binary protocols
(Bedrock's AWS Event Stream) that contain non-UTF-8 bytes which fail
JSON encoding.

Since fixture replay only uses the "b64" field, removed the redundant
"decoded" field. This is purely a fixture storage change - library
functionality unchanged.

Changes:
- Future streaming fixtures won't include "decoded" field
- Existing fixtures with "decoded" still work (backwards compatible)
- Fixes Jason.EncodeError when recording Bedrock streaming fixtures

Note: Bedrock uses AWS Event Stream binary protocol with headers/checksums,
unlike Anthropic's text-based SSE which encodes as valid UTF-8.

* Fix Bedrock extended thinking support

Bedrock provider was not calling Options.process like other providers
(Anthropic, Google, OpenAI), which meant pre_validate_options was never
invoked. This prevented reasoning_effort from being translated to the
thinking parameter needed by Bedrock's API.

Changes:
- Call Options.process in attach() to match other provider patterns
- Nest additional_model_request_fields under :provider_options (schema requirement)
- Use :thinking parameter (not :reasoning_config) per Bedrock API spec
- Use model.capabilities.reasoning instead of hardcoded model IDs
- Document that Bedrock returns reasoning_tokens as 0 (tokens included in output_tokens)

Extended thinking now works for all Claude models ... (continued)

Run Details

59 of 70 new or added lines in 4 files covered. (84.29%)

6 existing lines in 3 files now uncovered.

3705 of 7009 relevant lines covered (52.86%)

456.17 hits per line