fix: preserve non-ASCII text in provider request formatting

[kimnamu]

Thank you to the Strands maintainers for the excellent SDK — and for the recent #2653 which fixed the @tool decorator path. This PR addresses the provider-side counterpart of the same issue.

Fixes #2660

Problem

Provider request-formatting paths serialize tool-call arguments and tool-result {"json": ...} content blocks with json.dumps(), which defaults to ensure_ascii=True. Non-ASCII text (CJK, emoji) is therefore escaped to \uXXXX in the request sent to the model. For non-Latin scripts this inflates token usage (and cost) and hurts readability/debuggability.

This is the provider-side counterpart to the @tool decorator path fixed in #2653, and it matches what the SDK already does in telemetry/tracer.py and the session managers (ensure_ascii=False).

Reproduction (no network / API key needed)

from strands.models.openai import OpenAIModel

tool_result = {"toolUseId": "c1", "status": "success", "content": [{"json": {"city": "東京"}}]}
print(OpenAIModel.format_request_tool_message(tool_result)["content"])

	Before	After
OpenAI tool-result {"json": {"city": "東京"}}	{"city": "東京"} ❌	{"city": "東京"} ✅
OpenAI tool-call args {"query": "東京"}	{"query": "東京"} ❌	{"query": "東京"} ✅
Anthropic tool-result json	\uXXXX-escaped ❌	preserved ✅
Response-parsing paths (bedrock.py, ollama.py, gemini.py stream→internal)	unchanged	✅ unchanged (not request-facing)
Public API / signatures / output ordering	—	✅ unchanged

Change

json.dumps(x) → json.dumps(x, ensure_ascii=False) in the model-visible format_request_* paths only, across the affected providers (openai, anthropic, mistral, llamaapi, writer, llamacpp, ollama, openai_responses). Response→internal conversion paths are intentionally left untouched.

Tests

Added regression tests on the issue's primary paths (test_openai.py, test_anthropic.py). Verified they catch the bug:

• Source reverted (bug present) → 3 failed (e.g. assert '\u' not in '{"city": "東京"}')
• Source restored (fix) → 3 passed
• Full changed-provider suite → 398 passed, no regressions.
• ruff check / ruff format --check / mypy (changed files) all clean.

---

This contribution was prepared with the help of an AI agent (Claude Code); a human reviewed the change, rationale, and test results before submission.

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