feat: add DeepSeek V4 reasoning support - #28945
Open
andrew-manger wants to merge 1 commit into
Open
Conversation
DeepSeek V4 (deepseek-v4-flash/pro) supports thinking mode via:
- top-level reasoning_effort parameter (high/max)
- extra_body.thinking {type: enabled/disabled}
Per DeepSeek docs:
- low/medium/high map to 'high'
- xhigh maps to 'max'
- thinking defaults to enabled
This adds DeepSeek-specific handling in both:
1. ChatCompletionsTransport.build_kwargs() — main chat path
2. chat_completion_helpers summary path — compression/summary path
Hermes previously only supported reasoning for:
- Kimi (reasoning_effort + extra_body.thinking)
- TokenHub (reasoning_effort)
- LM Studio (reasoning_effort)
- OpenRouter Anthropic/Claude/OpenAI (extra_body.reasoning)
DeepSeek via direct API was silently dropping reasoning config.
Fixes: reasoning config now properly forwarded to DeepSeek V4 models.
alt-glitch
added
type/feature
This was referenced May 23, 2026
teknium1
reviewed
Jun 15, 2026
teknium1
left a comment
teknium1
left a comment
Contributor
There was a problem hiding this comment.
Thanks for the DeepSeek V4 reasoning work. The premise still has value, but current main has moved the active DeepSeek request-shaping path since this PR was opened.
Problems
- The main chat-path part is now wired to the wrong layer. Current main resolves registered providers at
agent/chat_completion_helpers.py:725-764and passesprovider_profileinto the transport;agent/transports/chat_completions.py:526-538then delegates provider-specific kwargs toprofile.build_api_kwargs_extras(). Forprovider: deepseek, the new legacyis_deepseekflag path in this PR would be bypassed. - The effort mapping requested here is not fully present on current main:
plugins/model-providers/deepseek/__init__.py:72-80still forwardslowandmediumunchanged, while this PR maps them to DeepSeek'shigh. - The summary-path addition appears to re-enable thinking for
/reasoning none: in the PR diff,agent/chat_completion_helpers.py:1018always sendssummary_extra_body["thinking"] = {"type": "enabled"}when the DeepSeek summary effort branch runs.
Suggested changes
- Salvage the mapping into
plugins/model-providers/deepseek/__init__.py::DeepSeekProfile.build_api_kwargs_extras(). - Add targeted tests for
high,xhigh,low/medium, and disabled reasoning, plus a summary-path kwargs test if summary support is retained.
Automated hermes-sweeper review.
| @@ -283,6 +284,25 @@ def build_kwargs( | |||
| _kimi_effort = _e | |||
Contributor
There was a problem hiding this comment.
On current main this lives in the bypassed legacy flag path for registered providers. provider: deepseek now routes through ProviderProfile, so this mapping should move into plugins/model-providers/deepseek/__init__.py::DeepSeekProfile.build_api_kwargs_extras().
| @@ -1009,6 +1016,9 @@ def handle_max_iterations(agent, messages: list, api_call_count: int) -> str: | |||
| summary_kwargs.update(agent._max_tokens_param(agent.max_tokens)) | |||
| if _lm_reasoning_effort is not None: | |||
| summary_kwargs["reasoning_effort"] = _lm_reasoning_effort | |||
Contributor
There was a problem hiding this comment.
This re-enables DeepSeek thinking during summaries even for /reasoning none: the branch is entered for a truthy reasoning_config, then always sends thinking.type=enabled. The disabled case needs to mirror the main request path.
teknium1
added
sweeper:risk-compatibility
This was referenced Jul 28, 2026
Open
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
DeepSeek V4 (deepseek-v4-flash / deepseek-v4-pro) added thinking mode with client-controllable reasoning effort. Hermes previously had no DeepSeek-specific reasoning handling, so reasoning config was silently dropped when using the direct DeepSeek API.
What DeepSeek V4 supports
Per DeepSeek API docs:
reasoning_efforttop-level parameter: "high" or "max"extra_body.thinking.type: "enabled" or "disabled"low/medium→high,xhigh→maxhigheffortChanges
agent/transports/chat_completions.pyis_deepseekflagreasoning_effortmapped to DeepSeek's levelsextra_body.thinkingtoggle (enabled/disabled)agent/chat_completion_helpers.py_is_deepseekdetectionis_deepseekthrough to transportextra_body.reasoningfor OpenRouter/Anthropic, andreasoning_effortfor LM Studio)Before / After
/reasoning high+ DeepSeek direct APIreasoning_effort: high+thinking: enabled/reasoning xhigh+ DeepSeek direct APIreasoning_effort: max+thinking: enabled/reasoning none+ DeepSeek direct APIthinking: disabledTest plan
deepseekwith modeldeepseek-v4-pro/reasoning high→ verifyreasoning_effort: highsent/reasoning xhigh→ verifyreasoning_effort: maxsent/reasoning none→ verifythinking: disabledsentRelated