[boogieLing]

Summary

Normalize non-string messages[*].content payloads before provider dispatch to improve DeepSeek/OpenAI-compatible request robustness.

Scope

• nanobot/providers/openai_compat_provider.py
• tests/providers/test_litellm_kwargs.py

Rationale

Some upstream tool/provider paths can produce non-string content blocks. This patch ensures stable provider input normalization without changing session/history behavior.

Validation

• uv run --extra dev pytest tests/providers/test_litellm_kwargs.py -q (passed)

Risk

Low. Provider payload normalization only; no session/memory lifecycle changes.

[chengyongru]

#3328

[Re-bin]

This is the right kind of fix.

The change is narrow and provider-scoped: DeepSeek gets strict string content normalization, while other OpenAI-compatible providers keep their existing block/list content behavior. That is the boundary I wanted to see here.

I pulled in the latest main and reviewed the three-dot diff. It stays limited to:

• nanobot/providers/openai_compat_provider.py
• tests/providers/test_litellm_kwargs.py

I also ran the relevant provider tests locally:

python -m pytest tests/providers/test_litellm_kwargs.py -q

Result: 58 passed.

From my side, this is clean, practical, and ready to merge.