fix: surface reasoning/thinking content as response instead of '(empty)'

[Sahil-SS9]

Fixes #58117

Description

When a thinking model (e.g. DeepSeek V4 Flash) returns a response where all
content is inside a reasoning/thinking block — or where content is empty
and only reasoning_content is populated — the agent exhausts its prefill
and empty-response retries and eventually sets final_response = "(empty)".

This causes:

• Blank responses — the user sees nothing despite the model having
  produced reasoning content
• Infinite cron heartbeat loops — cron jobs keep firing because the
  response is empty, wasting $30-80+ in API costs

Fix

When retries are exhausted and reasoning/thinking content IS available,
surface it as final_response instead of "(empty)". The reasoning
text (from reasoning_content, reasoning, or reasoning_details
fields) is the model's actual response — it was just delivered through the
thinking channel rather than visible content.

Verification

• Change is 9 lines in agent/conversation_loop.py
• No new dependencies
• Compile-check passes
• Only the reasoning-exhaustion exit path is affected — successful responses
  with visible content are unchanged

[alt-glitch]

This was generated by AI during triage.

Duplicate of #50397 — same fix in the reasoning-exhaustion exit path of agent/conversation_loop.py (surface reasoning_text as final_response instead of "(empty)" for reasoning-only models like DeepSeek V4 Flash). #50397 was opened earlier (2026-06-21) and is the more comprehensive version (also bumps thinking-prefill retries 2->3 and sets the persisted assistant message content), so it's canonical. Cross-linking so a maintainer can pick.

[tonydwb]

LGTM. Clean fix: surfaces reasoning/thinking content as response instead of '(empty)'. 1 file, 9 additions, well-scoped. No issues.

---

Reviewed by Hermes Agent

[teknium1]

Thanks for the focused fix. The premise is confirmed on current main: agent/conversation_loop.py:5177-5214 extracts reasoning but still terminates with (empty), and tests/run_agent/test_run_agent.py:4424-4444 reproduces that path.

Problems

• agent/conversation_loop.py:4895 leaves _turn_exit_reason as empty_response_exhausted. Current agent/turn_finalizer.py:338-366 treats a non-punctuated response of 24 characters or fewer under that reason as a partial fragment and appends the “⚠️ No reply” explainer. Thus a valid short fallback such as Yes would be delivered with a contradictory failure notice. The changed test fixture is 27 characters, so it misses this path.

Suggested changes

• Mark the reasoning fallback as a visible text completion, or explicitly exempt it from the partial-fragment explainer; add a short reasoning-only regression test.

This is an automated hermes-sweeper review.

[teknium1]

Please also change the terminal reason or exempt this fallback from the completion explainer. On current main, empty_response_exhausted plus a non-punctuated response of 24 characters or fewer makes agent/turn_finalizer.py:338-366 append “⚠️ No reply”; a valid reasoning fallback such as Yes would therefore receive a contradictory failure footer.

[Sahil-SS9]

Thanks @tonydwb for the review.

@alt-glitch — noted on the duplicate flag. #50397 was filed first and covers the same reasoning-exhaustion exit path. I checked #50397 and it's still open. Closing this one in favour of the earlier PR.

[Sahil-SS9]

Closing as duplicate of #50397 (filed earlier, same fix in the reasoning-exhaustion exit path). Thanks to @alt-glitch for flagging the overlap.