Summary

PR #15749 (merged) introduced a regression where assistant messages with both
and a valid field get their reasoning content overwritten
with an empty string when using DeepSeek or Kimi thinking mode.

Root Cause

In , #15749 reordered the logic so that the
check (step 2) runs before the field
promotion (step 3). This means:

1. A message has and
2. Step 2 sees + DeepSeek provider → injects → returns early
3. Step 3 (promote to ) is never reached
4. The valid reasoning content is lost

Expected Behavior

The field should be promoted to before any
empty-string fallback is applied. The empty-string injection should only happen
when there is no explicit AND no field to promote.

Reproduction

Proposed Fix

Reorder the steps in :

1. Preserve explicit
2. Promote field → (healthy path)
3. DeepSeek/Kimi fallback: inject for messages lacking both
4. Clean up non-string

This fix is already included in #15478 (rebased on latest main).

Related

• fix: DeepSeek/Kimi thinking mode requires reasoning_content on ALL assistant messages #15478 — fixes this regression + extends empty-string padding to all assistant messages (not just tool-call turns)
• [Bug]: DeepSeek V4 Flash Discord session poisoned when tool-call assistant lacks reasoning_content #15250 — original issue about missing reasoning_content causing HTTP 400
• fix(agent): ordering fix in _copy_reasoning_content_for_api — cross-provider reasoning isolation #15749 — the merged PR that introduced this regression

/cc @alt-glitch @OutThisLife