fix: extend DeepSeek reasoning_content fallback to text-only assistant messages

[brian-doherty]

Bug

When using DeepSeek in thinking mode (reasoning_effort: medium), text-only assistant responses (the final reply after a tool loop) could be persisted to the session DB without a reasoning_content field. On session resume, replaying that message to DeepSeek's API produces:

HTTP 400: The reasoning_content in the thinking mode must be passed back to the API.

This forced users to start a new session to recover.

Root Cause

_build_assistant_message (the creation path at run_agent.py:7763) only guaranteed the reasoning_content="" fallback for tool-call messages under DeepSeek thinking mode. Text-only responses fell through without setting reasoning_content at all, because:

1. The API response object may not have reasoning_content on every turn
2. The fallback only fired when msg["tool_calls"] was truthy AND _needs_deepseek_tool_reasoning() was True

Fix

Separates the Kimi (tool-call-only) and DeepSeek (all messages) fallback logic:

• Kimi: unchanged — reasoning_content="" only on tool-call messages
• DeepSeek: reasoning_content set on every assistant message, falling back to the content of the reasoning field or empty string

The replay safety net in _copy_reasoning_content_for_api (step 4) already catches this edge case, but storing correct data from creation time is cleaner and avoids relying on the replay-time catch-all.

Related

Refs #15250 (original DeepSeek tool-call reasoning_content fix), #15998 (this bug).

Test Plan

• Existing tests pass: pytest tests/run_agent/ -x -q
• Manual: DeepSeek thinking mode session with tool calls followed by text-only response, then session resume

[alt-glitch]

Part of the reasoning_content fix cluster — related to #15792, #15527, #15982, #16023. Verify this is not already addressed by one of the merged fixes.

[alt-glitch]

Part of the reasoning_content fix cluster.

[teknium1]

Closing as redundant — the DeepSeek reasoning_content thinking-mode 400 and cross-provider leak chain of issues is now fully covered on main:

• Direct provider detection + same-provider tool-call pad: commits 93a2d6b, d58b305, ad0ac89, 5ae6081
• All-assistant-messages pad rule: commit ad0ac89
• Ordering / cross-provider isolation: commit 9daa062 + regression guard 63bf7a2
• Cross-provider leak (MiniMax reasoning → DeepSeek): PR fix(agent): block cross-provider reasoning leak to DeepSeek/Kimi (#15748) #16500 (merging shortly)

21 regression tests in tests/run_agent/test_deepseek_reasoning_content_echo.py + 2 new tests for the cross-provider scenario exercise every known path. Thanks for the submission — appreciate the digging on this area.