Skip to content

fix(deepseek): preserve chat completions reasoning_content replay#16855

Closed
lsdsjy wants to merge 1 commit into
NousResearch:mainfrom
lsdsjy:fix/deepseek-v4-reasoning-content
Closed

fix(deepseek): preserve chat completions reasoning_content replay#16855
lsdsjy wants to merge 1 commit into
NousResearch:mainfrom
lsdsjy:fix/deepseek-v4-reasoning-content

Conversation

@lsdsjy

@lsdsjy lsdsjy commented Apr 28, 2026 •
edited
Loading

Copy link
Copy Markdown

Summary

  • Preserve DeepSeek reasoning_content from OpenAI SDK model_extra during response normalization and assistant message persistence.
  • Backfill reasoning_content for DeepSeek V4 thinking tool-call assistant turns so replay payloads include content, reasoning_content, and tool_calls.
  • Add regression coverage plus an opt-in live DeepSeek V4 thinking replay smoke test for both deepseek-v4-flash and deepseek-v4-pro.

Basis

DeepSeek official thinking-mode tool-call guidance requires full replay of reasoning_content after tool-call turns in all subsequent requests:

https://api-docs.deepseek.com/guides/thinking_mode#tool-calls:~:text=Please%20note%20that%2C%20unlike%20turns%20in%20thinking%20mode%20that%20do%20not%20involve%20tool%20calls%2C%20for%20turns%20that%20do%20perform%20tool%20calls%2C%20the%20reasoning_content%20must%20be%20fully%20passed%20back%20to%20the%20API%20in%20all%20subsequent%20requests.

Scope

This PR fixes the OpenAI-compatible Chat Completions path (/chat/completions) used by DeepSeek V4. Anthropic-format compatibility is intentionally out of scope here and is covered separately by #16781:

#16781

Testing

  • /Users/hfadmin/.hermes/hermes-agent/venv/bin/python -m py_compile run_agent.py agent/transports/chat_completions.py tests/run_agent/test_deepseek_reasoning_content_echo.py tests/run_agent/test_deepseek_v4_thinking_live.py tests/agent/transports/test_chat_completions.py
  • /Users/hfadmin/.hermes/hermes-agent/venv/bin/python -m pytest tests/run_agent/test_deepseek_reasoning_content_echo.py tests/agent/transports/test_chat_completions.py tests/run_agent/test_deepseek_v4_thinking_live.py -q (69 passed, 2 skipped)
  • HERMES_LIVE_TESTS=1 pytest tests/run_agent/test_deepseek_v4_thinking_live.py -q -s -n0 (2 passed, verified deepseek-v4-flash and deepseek-v4-pro against the live DeepSeek API)
  • git diff --check

@alt-glitch alt-glitch added type/bug Something isn't working P2 Medium — degraded but workaround exists comp/agent Core agent loop, run_agent.py, prompt builder provider/deepseek DeepSeek API labels Apr 28, 2026
@lsdsjy lsdsjy changed the title fix(deepseek): preserve v4 reasoning_content on replay fix(deepseek): preserve chat completions reasoning_content replay Apr 28, 2026
@teknium1 teknium1 mentioned this pull request Apr 30, 2026
Merged
teknium1 added a commit that referenced this pull request Apr 30, 2026
@teknium1 @season179
fix(agent): extend thinking-mode reasoning_content pad to Kimi/Moonshot
76edc40 
Builds on #16855 (@lsdsjy) which fixed DeepSeek v4 reasoning_content
replay via model_extra fallback + capturing tool_calls at method entry.
Kimi / Moonshot thinking mode enforces the same echo-back contract and
hits the same 400 when a tool-call turn is persisted without
reasoning_content.

- _build_assistant_message: pad branch now uses _needs_thinking_reasoning_pad()
  (DeepSeek OR Kimi) instead of _needs_deepseek_tool_reasoning() alone.
- Extract _needs_thinking_reasoning_pad() and reuse it in
  _copy_reasoning_content_for_api so both sites share one predicate.
- tests/run_agent/test_deepseek_reasoning_content_echo.py: add
  TestBuildAssistantMessagePadsStrictProviders parametrized over DeepSeek
  (attr=None, attr-absent), Kimi (attr=None), Moonshot (via base_url),
  and an OpenRouter negative control that must NOT pad. Proven to fail
  2/5 cases on Kimi/Moonshot without this change.
- scripts/release.py: add AUTHOR_MAP entries for lsdsjy and season179.

Refs #17400.

Co-authored-by: season179 <season.saw@gmail.com>
@teknium1

teknium1 commented Apr 30, 2026

Copy link
Copy Markdown
Contributor

Merged via PR #18045. Your commit (b9b9ee3e6) was cherry-picked onto current main with your authorship preserved in git log — the DeepSeek v4 + model_extra fallback + transport-layer normalizer fix is in exactly as you wrote it, plus the live replay smoke test. Thanks for the thorough fix.

Follow-up commit 76edc40ab extends the same pad to Kimi / Moonshot thinking mode (salvaged from #17489) since they enforce the same echo-back contract.

#18045

@teknium1 teknium1 mentioned this pull request Apr 30, 2026
Closed
13 tasks
donald131 pushed a commit to donald131/hermes-agent that referenced this pull request May 2, 2026
@teknium1 @season179
fix(agent): extend thinking-mode reasoning_content pad to Kimi/Moonshot
bf27542 
Builds on NousResearch#16855 (@lsdsjy) which fixed DeepSeek v4 reasoning_content
replay via model_extra fallback + capturing tool_calls at method entry.
Kimi / Moonshot thinking mode enforces the same echo-back contract and
hits the same 400 when a tool-call turn is persisted without
reasoning_content.

- _build_assistant_message: pad branch now uses _needs_thinking_reasoning_pad()
  (DeepSeek OR Kimi) instead of _needs_deepseek_tool_reasoning() alone.
- Extract _needs_thinking_reasoning_pad() and reuse it in
  _copy_reasoning_content_for_api so both sites share one predicate.
- tests/run_agent/test_deepseek_reasoning_content_echo.py: add
  TestBuildAssistantMessagePadsStrictProviders parametrized over DeepSeek
  (attr=None, attr-absent), Kimi (attr=None), Moonshot (via base_url),
  and an OpenRouter negative control that must NOT pad. Proven to fail
  2/5 cases on Kimi/Moonshot without this change.
- scripts/release.py: add AUTHOR_MAP entries for lsdsjy and season179.

Refs NousResearch#17400.

Co-authored-by: season179 <season.saw@gmail.com>
nickdlkk pushed a commit to nickdlkk/hermes-agent that referenced this pull request May 11, 2026
@teknium1 @season179 @nickdlkk
fix(agent): extend thinking-mode reasoning_content pad to Kimi/Moonshot
14364bd 
Builds on NousResearch#16855 (@lsdsjy) which fixed DeepSeek v4 reasoning_content
replay via model_extra fallback + capturing tool_calls at method entry.
Kimi / Moonshot thinking mode enforces the same echo-back contract and
hits the same 400 when a tool-call turn is persisted without
reasoning_content.

- _build_assistant_message: pad branch now uses _needs_thinking_reasoning_pad()
  (DeepSeek OR Kimi) instead of _needs_deepseek_tool_reasoning() alone.
- Extract _needs_thinking_reasoning_pad() and reuse it in
  _copy_reasoning_content_for_api so both sites share one predicate.
- tests/run_agent/test_deepseek_reasoning_content_echo.py: add
  TestBuildAssistantMessagePadsStrictProviders parametrized over DeepSeek
  (attr=None, attr-absent), Kimi (attr=None), Moonshot (via base_url),
  and an OpenRouter negative control that must NOT pad. Proven to fail
  2/5 cases on Kimi/Moonshot without this change.
- scripts/release.py: add AUTHOR_MAP entries for lsdsjy and season179.

Refs NousResearch#17400.

Co-authored-by: season179 <season.saw@gmail.com>
02356abc pushed a commit to 02356abc/hermes-agent that referenced this pull request May 14, 2026
@teknium1 @season179
fix(agent): extend thinking-mode reasoning_content pad to Kimi/Moonshot
c65d9ce 
Builds on NousResearch#16855 (@lsdsjy) which fixed DeepSeek v4 reasoning_content
replay via model_extra fallback + capturing tool_calls at method entry.
Kimi / Moonshot thinking mode enforces the same echo-back contract and
hits the same 400 when a tool-call turn is persisted without
reasoning_content.

- _build_assistant_message: pad branch now uses _needs_thinking_reasoning_pad()
  (DeepSeek OR Kimi) instead of _needs_deepseek_tool_reasoning() alone.
- Extract _needs_thinking_reasoning_pad() and reuse it in
  _copy_reasoning_content_for_api so both sites share one predicate.
- tests/run_agent/test_deepseek_reasoning_content_echo.py: add
  TestBuildAssistantMessagePadsStrictProviders parametrized over DeepSeek
  (attr=None, attr-absent), Kimi (attr=None), Moonshot (via base_url),
  and an OpenRouter negative control that must NOT pad. Proven to fail
  2/5 cases on Kimi/Moonshot without this change.
- scripts/release.py: add AUTHOR_MAP entries for lsdsjy and season179.

Refs NousResearch#17400.

Co-authored-by: season179 <season.saw@gmail.com>
jsboige pushed a commit to jsboige/hermes-agent that referenced this pull request May 14, 2026
@teknium1 @season179
fix(agent): extend thinking-mode reasoning_content pad to Kimi/Moonshot
cffff1d 
Builds on NousResearch#16855 (@lsdsjy) which fixed DeepSeek v4 reasoning_content
replay via model_extra fallback + capturing tool_calls at method entry.
Kimi / Moonshot thinking mode enforces the same echo-back contract and
hits the same 400 when a tool-call turn is persisted without
reasoning_content.

- _build_assistant_message: pad branch now uses _needs_thinking_reasoning_pad()
  (DeepSeek OR Kimi) instead of _needs_deepseek_tool_reasoning() alone.
- Extract _needs_thinking_reasoning_pad() and reuse it in
  _copy_reasoning_content_for_api so both sites share one predicate.
- tests/run_agent/test_deepseek_reasoning_content_echo.py: add
  TestBuildAssistantMessagePadsStrictProviders parametrized over DeepSeek
  (attr=None, attr-absent), Kimi (attr=None), Moonshot (via base_url),
  and an OpenRouter negative control that must NOT pad. Proven to fail
  2/5 cases on Kimi/Moonshot without this change.
- scripts/release.py: add AUTHOR_MAP entries for lsdsjy and season179.

Refs NousResearch#17400.

Co-authored-by: season179 <season.saw@gmail.com>
dannyJ848 pushed a commit to dannyJ848/hermes-agent that referenced this pull request May 17, 2026
@teknium1 @season179
fix(agent): extend thinking-mode reasoning_content pad to Kimi/Moonshot
a776523 
Builds on NousResearch#16855 (@lsdsjy) which fixed DeepSeek v4 reasoning_content
replay via model_extra fallback + capturing tool_calls at method entry.
Kimi / Moonshot thinking mode enforces the same echo-back contract and
hits the same 400 when a tool-call turn is persisted without
reasoning_content.

- _build_assistant_message: pad branch now uses _needs_thinking_reasoning_pad()
  (DeepSeek OR Kimi) instead of _needs_deepseek_tool_reasoning() alone.
- Extract _needs_thinking_reasoning_pad() and reuse it in
  _copy_reasoning_content_for_api so both sites share one predicate.
- tests/run_agent/test_deepseek_reasoning_content_echo.py: add
  TestBuildAssistantMessagePadsStrictProviders parametrized over DeepSeek
  (attr=None, attr-absent), Kimi (attr=None), Moonshot (via base_url),
  and an OpenRouter negative control that must NOT pad. Proven to fail
  2/5 cases on Kimi/Moonshot without this change.
- scripts/release.py: add AUTHOR_MAP entries for lsdsjy and season179.

Refs NousResearch#17400.

Co-authored-by: season179 <season.saw@gmail.com>
gweeteve pushed a commit to gweeteve/hermes-agent that referenced this pull request Jun 2, 2026
@teknium1 @season179
fix(agent): extend thinking-mode reasoning_content pad to Kimi/Moonshot
4b3e6ea 
Builds on NousResearch#16855 (@lsdsjy) which fixed DeepSeek v4 reasoning_content
replay via model_extra fallback + capturing tool_calls at method entry.
Kimi / Moonshot thinking mode enforces the same echo-back contract and
hits the same 400 when a tool-call turn is persisted without
reasoning_content.

- _build_assistant_message: pad branch now uses _needs_thinking_reasoning_pad()
  (DeepSeek OR Kimi) instead of _needs_deepseek_tool_reasoning() alone.
- Extract _needs_thinking_reasoning_pad() and reuse it in
  _copy_reasoning_content_for_api so both sites share one predicate.
- tests/run_agent/test_deepseek_reasoning_content_echo.py: add
  TestBuildAssistantMessagePadsStrictProviders parametrized over DeepSeek
  (attr=None, attr-absent), Kimi (attr=None), Moonshot (via base_url),
  and an OpenRouter negative control that must NOT pad. Proven to fail
  2/5 cases on Kimi/Moonshot without this change.
- scripts/release.py: add AUTHOR_MAP entries for lsdsjy and season179.

Refs NousResearch#17400.

Co-authored-by: season179 <season.saw@gmail.com>
Egavasyug pushed a commit to Egavasyug/hermes-agent that referenced this pull request Jun 10, 2026
@teknium1 @season179
fix(agent): extend thinking-mode reasoning_content pad to Kimi/Moonshot
cc8bdc7 
Builds on NousResearch#16855 (@lsdsjy) which fixed DeepSeek v4 reasoning_content
replay via model_extra fallback + capturing tool_calls at method entry.
Kimi / Moonshot thinking mode enforces the same echo-back contract and
hits the same 400 when a tool-call turn is persisted without
reasoning_content.

- _build_assistant_message: pad branch now uses _needs_thinking_reasoning_pad()
  (DeepSeek OR Kimi) instead of _needs_deepseek_tool_reasoning() alone.
- Extract _needs_thinking_reasoning_pad() and reuse it in
  _copy_reasoning_content_for_api so both sites share one predicate.
- tests/run_agent/test_deepseek_reasoning_content_echo.py: add
  TestBuildAssistantMessagePadsStrictProviders parametrized over DeepSeek
  (attr=None, attr-absent), Kimi (attr=None), Moonshot (via base_url),
  and an OpenRouter negative control that must NOT pad. Proven to fail
  2/5 cases on Kimi/Moonshot without this change.
- scripts/release.py: add AUTHOR_MAP entries for lsdsjy and season179.

Refs NousResearch#17400.

Co-authored-by: season179 <season.saw@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

comp/agent Core agent loop, run_agent.py, prompt builder P2 Medium — degraded but workaround exists provider/deepseek DeepSeek API type/bug Something isn't working

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@lsdsjy @teknium1 @alt-glitch