Skip to content

fix: extract DeepSeek reasoning_content from model_extra#14973

Closed
HiddenPuppy wants to merge 1 commit into
NousResearch:mainfrom
HiddenPuppy:fix/openrouter-tools-400
Closed

fix: extract DeepSeek reasoning_content from model_extra#14973
HiddenPuppy wants to merge 1 commit into
NousResearch:mainfrom
HiddenPuppy:fix/openrouter-tools-400

Conversation

@HiddenPuppy

@HiddenPuppy HiddenPuppy commented Apr 24, 2026

Copy link
Copy Markdown
Contributor

Problem

DeepSeek V4 Flash has thinking mode enabled by default. When the model makes a tool call, the API returns a reasoning_content field in the response.

However, OpenAI SDK < 1.60 doesn't declare reasoning_content as a ChatCompletionMessage field. It ends up in Pydantic's model_extra instead. Without this fix, the reasoning_content is lost and subsequent requests fail with:

Changes

  • _extract_reasoning() now checks model_extra first, then falls back to the direct attribute for backward compatibility with newer SDK versions.
  • Added tests to verify extraction from both model_extra and direct attributes.

Fixes

Test Plan

============================= test session starts ==============================
platform linux -- Python 3.11.6, pytest-9.0.3, pluggy-1.6.0 -- /usr/bin/python3
cachedir: .pytest_cache
rootdir: /tmp/hermes-agent
configfile: pyproject.toml
plugins: anyio-4.13.0, asyncio-1.3.0, xdist-3.8.0
asyncio: mode=Mode.STRICT, debug=False, asyncio_default_fixture_loop_scope=None, asyncio_default_test_loop_scope=function
created: 2/2 workers
2 workers [6 items]

scheduling tests via LoadScheduling

tests/run_agent/test_deepseek_reasoning_content.py::TestExtractReasoning::test_extract_from_model_extra
tests/run_agent/test_deepseek_reasoning_content.py::TestExtractReasoning::test_model_extra_takes_precedence_over_attribute
[gw1] [ 16%] PASSED tests/run_agent/test_deepseek_reasoning_content.py::TestExtractReasoning::test_model_extra_takes_precedence_over_attribute
[gw0] [ 33%] PASSED tests/run_agent/test_deepseek_reasoning_content.py::TestExtractReasoning::test_extract_from_model_extra
tests/run_agent/test_deepseek_reasoning_content.py::TestExtractReasoning::test_extract_from_direct_attribute
tests/run_agent/test_deepseek_reasoning_content.py::TestExtractReasoning::test_no_reasoning_returns_none
[gw0] [ 50%] PASSED tests/run_agent/test_deepseek_reasoning_content.py::TestExtractReasoning::test_extract_from_direct_attribute
[gw1] [ 66%] PASSED tests/run_agent/test_deepseek_reasoning_content.py::TestExtractReasoning::test_no_reasoning_returns_none
tests/run_agent/test_deepseek_reasoning_content.py::TestExtractReasoning::test_none_model_extra
tests/run_agent/test_deepseek_reasoning_content.py::TestExtractReasoning::test_empty_model_extra
[gw1] [ 83%] PASSED tests/run_agent/test_deepseek_reasoning_content.py::TestExtractReasoning::test_empty_model_extra
[gw0] [100%] PASSED tests/run_agent/test_deepseek_reasoning_content.py::TestExtractReasoning::test_none_model_extra

============================== 6 passed in 3.60s ===============================
============================= test session starts ==============================
platform linux -- Python 3.11.6, pytest-9.0.3, pluggy-1.6.0 -- /usr/bin/python3
cachedir: .pytest_cache
rootdir: /tmp/hermes-agent
configfile: pyproject.toml
plugins: anyio-4.13.0, asyncio-1.3.0, xdist-3.8.0
asyncio: mode=Mode.STRICT, debug=False, asyncio_default_fixture_loop_scope=None, asyncio_default_test_loop_scope=function
created: 2/2 workers
2 workers [39 items]

scheduling tests via LoadScheduling

tests/agent/transports/test_chat_completions.py::TestChatCompletionsBasic::test_convert_messages_strips_codex_fields
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBasic::test_api_mode
[gw1] [ 2%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBasic::test_convert_messages_strips_codex_fields
[gw0] [ 5%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBasic::test_api_mode
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBasic::test_registered
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_basic_kwargs
[gw0] [ 7%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBasic::test_registered
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBasic::test_convert_tools_identity
[gw0] [ 10%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBasic::test_convert_tools_identity
[gw1] [ 12%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_basic_kwargs
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_developer_role_swap
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBasic::test_convert_messages_no_codex_leaks
[gw1] [ 15%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_developer_role_swap
[gw0] [ 17%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBasic::test_convert_messages_no_codex_leaks
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_tools_included
[gw0] [ 20%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_tools_included
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_openrouter_provider_prefs
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_no_developer_swap_for_non_gpt5
[gw0] [ 23%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_openrouter_provider_prefs
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_nous_tags
[gw1] [ 25%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_no_developer_swap_for_non_gpt5
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_ollama_num_ctx
[gw0] [ 28%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_nous_tags
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_reasoning_default
[gw1] [ 30%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_ollama_num_ctx
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_custom_think_false
[gw0] [ 33%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_reasoning_default
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_nous_omits_disabled_reasoning
[gw1] [ 35%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_custom_think_false
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_max_tokens_with_fn
[gw1] [ 38%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_max_tokens_with_fn
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_ephemeral_overrides_max_tokens
[gw1] [ 41%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_ephemeral_overrides_max_tokens
[gw0] [ 43%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_nous_omits_disabled_reasoning
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_nvidia_default_max_tokens
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_fixed_temperature
[gw1] [ 46%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_fixed_temperature
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_omit_temperature
[gw0] [ 48%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_nvidia_default_max_tokens
[gw1] [ 51%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_omit_temperature
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_qwen_default_max_tokens
tests/agent/transports/test_chat_completions.py::TestChatCompletionsKimi::test_kimi_max_tokens_default
[gw1] [ 53%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsKimi::test_kimi_max_tokens_default
tests/agent/transports/test_chat_completions.py::TestChatCompletionsKimi::test_kimi_reasoning_effort_top_level
[gw1] [ 56%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsKimi::test_kimi_reasoning_effort_top_level
[gw0] [ 58%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_qwen_default_max_tokens
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_anthropic_max_output_for_claude_on_aggregator
tests/agent/transports/test_chat_completions.py::TestChatCompletionsKimi::test_kimi_reasoning_effort_omitted_when_thinking_disabled
[gw0] [ 61%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_anthropic_max_output_for_claude_on_aggregator
[gw1] [ 64%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsKimi::test_kimi_reasoning_effort_omitted_when_thinking_disabled
tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_request_overrides_last
tests/agent/transports/test_chat_completions.py::TestChatCompletionsKimi::test_kimi_thinking_enabled_extra_body
[gw0] [ 66%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsBuildKwargs::test_request_overrides_last
[gw1] [ 69%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsKimi::test_kimi_thinking_enabled_extra_body
tests/agent/transports/test_chat_completions.py::TestChatCompletionsKimi::test_kimi_thinking_disabled_extra_body
[gw1] [ 71%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsKimi::test_kimi_thinking_disabled_extra_body
tests/agent/transports/test_chat_completions.py::TestChatCompletionsValidate::test_none
[gw0] [ 74%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsValidate::test_none
tests/agent/transports/test_chat_completions.py::TestChatCompletionsValidate::test_empty_choices
tests/agent/transports/test_chat_completions.py::TestChatCompletionsValidate::test_no_choices
[gw1] [ 76%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsValidate::test_empty_choices
[gw0] [ 79%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsValidate::test_no_choices
tests/agent/transports/test_chat_completions.py::TestChatCompletionsValidate::test_valid
[gw1] [ 82%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsValidate::test_valid
tests/agent/transports/test_chat_completions.py::TestChatCompletionsNormalize::test_text_response
[gw0] [ 84%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsNormalize::test_text_response
tests/agent/transports/test_chat_completions.py::TestChatCompletionsNormalize::test_tool_call_extra_content_preserved
[gw0] [ 87%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsNormalize::test_tool_call_extra_content_preserved
tests/agent/transports/test_chat_completions.py::TestChatCompletionsNormalize::test_tool_call_response
[gw1] [ 89%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsNormalize::test_tool_call_response
tests/agent/transports/test_chat_completions.py::TestChatCompletionsCacheStats::test_no_usage
[gw1] [ 92%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsCacheStats::test_no_usage
tests/agent/transports/test_chat_completions.py::TestChatCompletionsNormalize::test_reasoning_content_preserved_separately
[gw0] [ 94%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsNormalize::test_reasoning_content_preserved_separately
tests/agent/transports/test_chat_completions.py::TestChatCompletionsCacheStats::test_no_details
[gw1] [ 97%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsCacheStats::test_no_details
tests/agent/transports/test_chat_completions.py::TestChatCompletionsCacheStats::test_with_cache
[gw0] [100%] PASSED tests/agent/transports/test_chat_completions.py::TestChatCompletionsCacheStats::test_with_cache

============================== 39 passed in 0.67s ==============================

All tests pass.

fix: extract DeepSeek reasoning_content from model_extra (NousResearc…
b460cd0 
…h#14938)

DeepSeek V4 Flash returns reasoning_content in API responses, but OpenAI
SDK < 1.60 doesn't declare it as a ChatCompletionMessage field. It ends up
in Pydantic's model_extra instead.

Without this fix, the reasoning_content is lost and subsequent requests fail
with HTTP 400: The reasoning_content in the thinking mode must be passed back
to the API.

Changes:
- _extract_reasoning() now checks model_extra first, then falls back to the
direct attribute for backward compatibility with newer SDK versions.
- Added tests to verify extraction from both model_extra and direct attributes.

Fixes NousResearch#14938, NousResearch#14933
@alt-glitch alt-glitch added type/bug Something isn't working P1 High — major feature broken, no workaround comp/agent Core agent loop, run_agent.py, prompt builder provider/deepseek DeepSeek API labels Apr 24, 2026
@CompaCompu CompaCompu mentioned this pull request Apr 24, 2026
Closed
@huacao59109

huacao59109 commented Apr 24, 2026

Copy link
Copy Markdown

I hit a related but slightly different reproduction in a real scheduled Hermes run, and wanted to add the evidence here in case #14973 only covers the "DeepSeek generated reasoning_content but Hermes failed to extract it from model_extra" path.

Scenario

  • Primary model/provider: openai-codex / gpt-5.5
  • Fallback provider/model: custom OpenAI-compatible provider deepseek-v4 / deepseek-v4-flash
  • Fallback endpoint: https://api.deepseek.com/v1
  • Execution path: cron job (daily-github-trending-brief) that uses tools (web_extract, browser tools, terminal/lark-cli) and then auto-delivers the final response.

Timeline / logs

The primary model hit rate limits earlier in the run window, and the cron run activated fallback successfully:

2026-04-24 19:00:12,619 INFO [cron_f8de41fd3ff8_20260424_190002] root: Fallback activated: gpt-5.5 → deepseek-v4-flash (deepseek-v4)

But the fallback run then failed before producing a final response:

2026-04-24 19:01:16,685 ERROR [cron_f8de41fd3ff8_20260424_190002] root: Non-retryable client error: Error code: 400 - {'error': {'message': 'The `reasoning_content` in the thinking mode must be passed back to the API.', 'type': 'invalid_request_error', 'param': None, 'code': 'invalid_request_error'}}

The saved cron output confirms that delivery did not fail; no final assistant response was generated:

## Response

(No response generated)

Request-dump shape

The failing request dump was a DeepSeek chat-completions request:

POST https://api.deepseek.com/v1/chat/completions
model: deepseek-v4-flash

The request history contained assistant messages from the pre-fallback execution with tool calls and reasoning_content, e.g. multiple messages shaped like:

{
  "role": "assistant",
  "content": "",
  "tool_calls": [...],
  "reasoning_content": "..."
}

There were also tool result messages after those assistant tool calls. In the dump I checked, 10+ assistant messages had reasoning_content; most were tool-call turns produced before/around the fallback.

Why this may be a separate edge case

#14973 appears to fix the case where DeepSeek itself returns reasoning_content in model_extra, Hermes fails to extract it, and the next DeepSeek request omits it.

This cron failure seems like a cross-provider fallback/migration case:

  1. The turn starts on openai-codex / gpt-5.5.
  2. Hermes records assistant/tool-call history, including reasoning fields from the original provider path.
  3. Hermes switches in-place to deepseek-v4-flash after the primary fails.
  4. DeepSeek receives a mixed-provider tool-call history and rejects the thinking-mode replay with the same reasoning_content error.

I also saw a similar non-cron/gateway case today after switching from the primary model to DeepSeek; the immediate workaround was setting reasoning to none, which suggests this is not limited to cron delivery.

Expected behavior

Fallback to DeepSeek should either:

  • preserve and replay DeepSeek-native reasoning_content correctly when the previous turns were actually generated by DeepSeek; or
  • when migrating an existing conversation/tool-call history from a different provider to DeepSeek, sanitize provider-specific reasoning replay fields and/or disable DeepSeek thinking for that migrated turn, so fallback can complete instead of failing with 400.

In short: please consider testing both paths:

  1. DeepSeek → tool call → DeepSeek continuation (the model_extra extraction case fixed by this PR)
  2. OpenAI/Codex → tool call(s) → fallback to DeepSeek continuation (cross-provider fallback case)

Happy to provide more details from the local request dump if useful, but I avoided pasting the full dump because it includes large tool outputs and deployment-specific prompt/config text.

@ukint-vs ukint-vs mentioned this pull request Apr 24, 2026
Closed
@Zjianru Zjianru mentioned this pull request Apr 25, 2026
Closed
@alt-glitch alt-glitch mentioned this pull request Apr 25, 2026
Closed
@teknium1

teknium1 commented Apr 27, 2026

Copy link
Copy Markdown
Contributor

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

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.

@teknium1 teknium1 closed this Apr 27, 2026
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 P1 High — major feature broken, no workaround 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.

DeepSeek V4 Flash Toolcall Fails with reasoning_content Error [BUG] DeepSeek V4 thinking mode fails with reasoning_content error

4 participants

@HiddenPuppy @huacao59109 @teknium1 @alt-glitch