Summary

• _filter_restored_tool_context() stripped toolUse/toolResult blocks from restored session history but left reasoningContent (extended thinking) blocks intact.
• Bedrock rejects an assistant message where reasoningContent blocks are present without their companion toolUse blocks, raising a ValidationException.
• Fix: add "reasoningContent" not in content to the filter predicate so thinking blocks are stripped together with the tool calls they preceded.

Root cause

When extended thinking is enabled, an assistant turn produces a message like:

{
  "role": "assistant",
  "content": [
    {"reasoningContent": {"reasoningText": {"text": "…"}}},
    {"toolUse": {"toolUseId": "t1", "name": "my_tool", "input": {}}}
  ]
}

reasoningContent and toolUse are semantically coupled — the Bedrock API requires them to appear together. filter_restored_tool_context was removing toolUse while keeping reasoningContent, producing a partial message that the API refuses.

Test plan

• test_strips_reasoning_content_alongside_tool_use — assistant message with reasoningContent+toolUse → both stripped, only plain-text assistant message survives
• test_message_with_only_reasoning_and_tool_use_is_dropped_entirely — message becomes empty after stripping → excluded from output
• test_text_alongside_reasoning_and_tool_use_is_preserved — text content survives after stripping reasoning+toolUse
• test_messages_without_tool_context_are_unchanged — plain text messages pass through unchanged (no regression)
• test_existing_tool_use_filtering_still_works — original toolUse/toolResult behaviour preserved

5 passed in 0.XXs

Fixes #621