fix: joined tool_results to prevent context lost

[poshinchen]

Description

Fixes a bug where multi-block toolResult.content was silently dropped to the first block, causing FaithfulnessEvaluator (and other evaluators) to flag values from content[1+] as hallucinated.

Root cause: Three sites in StrandsInMemorySessionMapper and two in CloudWatchSessionMapper read content[0] (or response[0]) only:

• StrandsInMemorySessionMapper._process_tool_results (legacy convention)
• StrandsInMemorySessionMapper._convert_inference_messages (latest convention — had a ## To-do comment)
• StrandsInMemorySessionMapper._convert_tool_execution_span (latest convention)
• CloudWatchSessionMapper._extract_tool_results / _process_tool_results (via _extract_tool_result_text)

Fix: Two shared helpers in mappers/utils.py, used by both mappers:

• serialize_tool_result_block(block) — serializes a single Bedrock-style content block. Handles text, json (via json.dumps), and image/document/video (placeholder marker so the judge knows non-text data exists).
• join_tool_result_content(content) — joins all blocks with \n, filtering empties.

This addresses both the reported text-block bug and the latent same-shape bug for json blocks (common via Strands' @tool decorator returning dicts).

Related Issues

Fixes #235

Ref: aws/agentcore-cli#1393

Documentation PR

N/A

Type of Change

Bug fix

Testing

• 4 new regression tests in test_strands_in_memory_mapper.py covering: legacy multi-text, legacy text+json, latest-convention multi-text inference, latest-convention multi-text tool execution span.

• 1 new regression test in test_cloudwatch_session_mapper.py for multi-block tool result reaching the ToolExecutionSpan.

• All 172 existing mapper tests pass unchanged.

• I ran hatch run prepare

Checklist

• I have read the CONTRIBUTING document
• I have added any necessary tests that prove my fix is effective or my feature works
• I have updated the documentation accordingly
• I have added an appropriate example to the documentation to outline the feature, or no new docs are needed
• My changes generate no new warnings
• Any dependent changes have been merged and published

---

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

[github-actions]

Issue: serialize_tool_result_block and join_tool_result_content have multiple code paths (text, json, image/document/video, non-dict, None, str, list) but no direct unit tests in test_utils.py. The existing regression tests only exercise the text and json paths through the mapper integration layer.

Suggestion: Add focused unit tests in tests/strands_evals/mappers/test_utils.py for these functions covering edge cases:

• serialize_tool_result_block with: None, non-dict, {"text": ""}, {"json": {...}}, {"json": <unserializable>}, {"image": ...}, {"document": ...}, {"video": ...}, empty dict
• join_tool_result_content with: None, [], "", a plain string, a list with mixed block types, a non-list/non-str value

[github-actions]

Assessment: Approve

Clean, well-scoped bug fix that correctly consolidates duplicated content[0]-only logic into shared helpers. The approach is DRY, handles edge cases thoughtfully, and the regression tests prove the fix works end-to-end.

Review Details

• Testing: The new helpers have good indirect coverage through mapper tests, but would benefit from direct unit tests in test_utils.py to lock down each branch (image/document/video placeholders, json serialization fallback, non-dict inputs, etc.)
• Code quality: Implementation is clean — good use of early returns, proper fallback handling for json.dumps, and sensible placeholder markers for binary content types.

Nice consolidation that eliminates the ## To-do comment and fixes both text and json block handling in one pass.

[poshinchen]

closing it since #240 is preferred.