Fix deepseek-reasoner multi-turn requests missing reasoning_content

[iuhoay]

Summary

DeepSeek's thinking-capable models (deepseek-reasoner, the newer deepseek-v4-flash / deepseek-v4-pro, and presumably anything else DeepSeek ships with thinking mode) reject multi-turn requests when any prior assistant message is missing a reasoning_content key:

The reasoning_content in the thinking mode must be passed back to the API.

The inherited OpenAI format_thinking (lib/ruby_llm/providers/openai/chat.rb:111-124) returns {} when msg.thinking is nil and omits both reasoning / reasoning_content when thinking.text is empty. Any assistant turn that emitted only tool calls — or that simply didn't capture thinking text — therefore breaks the next request.

This PR overrides format_thinking on RubyLLM::Providers::DeepSeek::Chat to always emit reasoning_content on assistant messages (empty string when no thinking text was captured), preserving the existing reasoning / reasoning_signature semantics for everything else.

The override is provider-level (not gated on model id), so it applies to every DeepSeek model. For non-thinking models like deepseek-chat the API ignores the empty key.

Test plan

• New spec at spec/ruby_llm/providers/deep_seek/chat_spec.rb covers: thinking text + signature present, no thinking captured (the regression case), and non-assistant messages
• bundle exec rspec spec/ruby_llm/providers/deep_seek/chat_spec.rb — 3 examples, 0 failures
• bundle exec rubocop lib/ruby_llm/providers/deepseek/chat.rb spec/ruby_llm/providers/deep_seek/chat_spec.rb — no offenses
• Reviewer to confirm the existing deepseek-chat VCR-backed flows still pass (cassettes were recorded before this change; the empty reasoning_content should be ignored, but worth a re-record if the live API tightens validation)

🤖 Generated with Claude Code

[ashishbista]

+1

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 81.76%. Comparing base (a5f3dbd) to head (c09cfc4).
⚠️ Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #749      +/-   ##
==========================================
+ Coverage   81.67%   81.76%   +0.08%     
==========================================
  Files         169      169              
  Lines        7641     7684      +43     
  Branches     1269     1281      +12     
==========================================
+ Hits         6241     6283      +42     
+ Misses        891      889       -2     
- Partials      509      512       +3     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.