bug(strands): normalize public reasoning and keep Agent output final-only #282

Problem

Strands exposes provider-public reasoning as reasoningContent.reasoningText.text. The instrumentation currently passes the raw SDK block into GenAI messages:

{"reasoningContent":{"reasoningText":{"text":"synthetic plan"}}}

This is missing the shared message-part discriminator. It should become:

{"type":"reasoning","content":"synthetic plan"}

The Agent invocation also reuses the model-output converter, so its final output includes the final model's reasoning alongside the answer. Agent output should describe the final result; intermediate reasoning, tool plans and tool results belong to the model/tool spans.

Reproduction and evidence

Using Strands 1.60.0 with the real Agent.invoke_async / Agent.stream_async loop, a deterministic provider, reasoning deltas and actual tool dispatch:

• Two user turns per session, multiple model decisions and concurrent sessions reproduce both defects.
• A complex SDK-backed application produces 18 raw reasoning blocks on 18 LLM spans and reasoning in all 6 Agent outputs.
• Streaming and non-streaming regression assertions fail before the fix. Changing only reasoning conversion still leaves the Agent boundary assertion failing.
• The application's final answers, SDK messages and tool results are otherwise correct.

These defects predate the separate stream lifecycle fix in #278.

Expected behavior

• Normalize public reasoning text in model inputs and outputs through the shared Reasoning type.
• Exclude reasoning signatures, encrypted/redacted reasoning and unavailable text from telemetry.
• Keep Agent output limited to final result parts, preserving final multimodal parts.
• Preserve original SDK responses, stream results, tools, usage and context behavior.
• Apply the same message semantics to span and structured log-event capture.

A focused follow-up fix and regression tests are being prepared.