fix(transport): emit deepseek-v4 thinking.type and reasoning_effort on non-OpenRouter routes by Skyline10124 · Pull Request #16614 · NousResearch/hermes-agent

Skyline10124 · 2026-04-27T15:19:34Z

What does this PR do?

Add DeepSeek V4 thinking mode support that works through any route (direct api.deepseek.com, opencode-go, or other OpenAI-compatible relays), not just OpenRouter.

Currently _supports_reasoning_extra_body() gates on base_url containing "openrouter", which silently drops reasoning_config for direct DeepSeek connections and opencode-go. This PR makes reasoning_effort and thinking.type actually reach the API.

Related Issue

Fixes #15717

Related to #14958, #15251 — those PRs correct the detection gate but stop at run_agent.py. This PR completes the pipeline end-to-end.

Type of Change

🐛 Bug fix (reasoning_effort was silently dropped for non-OpenRouter routes)
✨ New feature (DeepSeek V4 thinking mode via any route)

Changes Made

run_agent.py — _supports_reasoning_extra_body() (1 line)

Detect DeepSeek V4 models by name prefix (deepseek-v4), so routes like opencode-go (opencode.ai/zen/go/v1) and any future relay can enable thinking.

agent/transports/chat_completions.py — build_kwargs() (28 lines)

New is_deepseek_v4 branch, modelled after the existing is_kimi handling
Top-level reasoning_effort with effort mapping: xhigh → max, others → high
extra_body["thinking"] = {"type": "enabled"/"disabled"} — DeepSeek-native format
Exclude DeepSeek V4 from the generic extra_body["reasoning"] path to avoid format conflict

tests/agent/transports/test_chat_completions.py — 7 new tests

reasoning_effort top-level (high, xhigh→max, disabled→omitted)
extra_body.thinking enabled/disabled
No stray extra_body.reasoning for V4
deepseek-chat (V2/V3) untouched — regression guard

How to Test

Configure a DeepSeek V4 model without OpenRouter (e.g. provider: opencode-go, base_url: https://opencode.ai/zen/go/v1, reasoning_effort: xhigh)
Start a session and ask a reasoning-intensive question
Verify the API request includes reasoning_effort: "max" and extra_body: {"thinking": {"type": "enabled"}}

Or run the transport-level unit tests:

python -m pytest tests/agent/transports/test_chat_completions.py -k "deepseek" -v

Checklist

Code

I've read the Contributing Guide
My commit messages follow Conventional Commits
I searched for existing PRs (feat(deepseek): plumb reasoning_effort and thinking toggle to V4 API #14958, feat(deepseek): add thinking.type + reasoning_effort mapping for DeepSeek API #15251, feat: add DeepSeek-V4 thinking mode via unified thinking_mode parameter #15577) — this PR takes a different approach (model-name gating + transport-level handling, no new config parameter)
My PR contains only changes related to this fix
I've run pytest tests/agent/transports/test_chat_completions.py — 48/48 passing (41 existing + 7 new)
I've added tests for my changes
I've tested on Arch Linux (Hyprland)

Documentation & Housekeeping

I've updated relevant documentation — or N/A
I've updated cli-config.yaml.example — or N/A
I've considered cross-platform impact — or N/A

Why this approach

Existing PRs (#14958, #15251) add is_deepseek detection in run_agent.py but don't touch build_kwargs() — so the transport still emits OpenRouter-style extra_body["reasoning"] format. #15577 introduces a unified thinking_mode parameter framework.

This PR takes the minimal path:

No new config parameter — reuses the existing reasoning_config flow
Model-name-based gating — works for opencode-go, direct API, and any future relay
Scoped to V4 only — deepseek-v4* prefix match; deepseek-chat / deepseek-reasoner (legacy) are untouched
Follows existing patterns — the code structure mirrors is_kimi handling, keeping the transport predictable

Copilot

Pull request overview

Extends the agent → transport plumbing so DeepSeek V4 “thinking mode” parameters (reasoning_effort and extra_body.thinking.type) are emitted based on the model name (e.g., deepseek-v4-*), enabling non-OpenRouter routes (direct DeepSeek API, OpenAI-compatible relays) to receive the correct request shape.

Changes:

Update run_agent.py reasoning-extra-body gating to treat deepseek-v4* as reasoning-capable.
Update ChatCompletionsTransport.build_kwargs() to emit DeepSeek V4’s native reasoning_effort + extra_body.thinking and to avoid the generic extra_body["reasoning"] path for V4.
Add transport-level unit tests covering DeepSeek V4 request shaping and guarding legacy deepseek-chat behavior.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
tests/agent/transports/test_chat_completions.py	Adds DeepSeek V4-focused unit tests for `reasoning_effort` and `extra_body.thinking` behavior.
run_agent.py	Adjusts `_supports_reasoning_extra_body()` to return true for `deepseek-v4*` models.
agent/transports/chat_completions.py	Adds DeepSeek V4 request-shaping branch and excludes V4 from generic `extra_body.reasoning`.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-27T15:23:43Z

+        # DeepSeek V4 series: supports reasoning_effort + thinking via any route
+        if self.model and self.model.lower().startswith("deepseek-v4"):
+            return True
+


_supports_reasoning_extra_body() is also used in _handle_max_iterations() to decide whether to send extra_body["reasoning"] (OpenRouter-style). Returning True for deepseek-v4* will make iteration-limit summary requests include extra_body.reasoning, which conflicts with the DeepSeek V4 contract (should use top-level reasoning_effort + extra_body.thinking). Suggest keeping this method scoped to the generic extra_body.reasoning format and instead handle DeepSeek V4 summary kwargs via the chat_completions transport (or add an explicit DeepSeek V4 branch in the summary path).

Copilot · 2026-04-27T15:23:44Z

@@ -188,6 +188,7 @@ def build_kwargs(
        anthropic_max_out = params.get("anthropic_max_output")
        is_nvidia_nim = params.get("is_nvidia_nim", False)
        is_kimi = params.get("is_kimi", False)
+        is_deepseek_v4 = params.get("model_lower", "").startswith("deepseek-v4")


DeepSeek V4 detection uses params.get("model_lower", ""), which means the DeepSeek V4 branch silently won’t run if a caller relies on the existing fallback model_lower = params.get("model_lower", (model or "").lower()) (i.e., doesn’t pass model_lower). Use the already-computed local model_lower (or (model or "").lower()) to keep detection consistent across call sites.

Suggested change

is_deepseek_v4 = params.get("model_lower", "").startswith("deepseek-v4")

is_deepseek_v4 = model_lower.startswith("deepseek-v4")

Copilot · 2026-04-27T15:23:44Z

@@ -240,8 +256,19 @@ def build_kwargs(
                "type": "enabled" if _kimi_thinking_enabled else "disabled",
            }

+        # DeepSeek V4: extra_body.thinking
+        if is_deepseek_v4:
+            _ds_thinking_off = bool(
+                reasoning_config
+                and isinstance(reasoning_config, dict)
+                and reasoning_config.get("enabled") is False
+            )
+            extra_body["thinking"] = {
+                "type": "enabled" if not _ds_thinking_off else "disabled",
+            }


The DeepSeek V4 block repeats the same _ds_thinking_off computation in two places (reasoning_effort + extra_body.thinking). Consider computing it once and reusing it to reduce duplication and the chance of future drift between the two branches.

Copilot

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-27T15:48:28Z

+        _ds_thinking_off = bool(
+            reasoning_config
+            and isinstance(reasoning_config, dict)
+            and reasoning_config.get("enabled") is False


_ds_thinking_off only checks reasoning_config['enabled'] is False, but other parts of the codebase pass reasoning_config={"effort": "none"} to disable thinking (e.g., batch_runner's --reasoning_disabled). For DeepSeek V4 this currently still sends reasoning_effort and sets extra_body.thinking.type to enabled. Consider treating effort == "none" (case/whitespace-normalized) as thinking disabled here as well, and omit reasoning_effort in that case.

Suggested change

and reasoning_config.get("enabled") is False

and (

reasoning_config.get("enabled") is False

or (reasoning_config.get("effort") or "").strip().lower() == "none"

)

Copilot · 2026-04-27T15:48:29Z

+                    _ds_thinking_off = (
+                        self.reasoning_config is not None
+                        and isinstance(self.reasoning_config, dict)
+                        and self.reasoning_config.get("enabled") is False
+                    )
+                    summary_extra_body["thinking"] = {
+                        "type": "enabled" if not _ds_thinking_off else "disabled",
                    }


DeepSeek V4 summary requests treat thinking as disabled only when reasoning_config.enabled is False, but callers may also disable thinking via reasoning_config={"effort": "none"} (used elsewhere in the codebase). With the current check, V4 summary calls will still set extra_body.thinking.type to enabled even when the user requested no thinking. Update _ds_thinking_off to also consider effort == "none" as disabled (normalize case/whitespace).

Copilot · 2026-04-27T15:48:29Z

+    def test_deepseek_v4_reasoning_effort_omitted_when_thinking_disabled(self, transport):
+        kw = transport.build_kwargs(
+            model="deepseek-v4-pro", messages=[{"role": "user", "content": "Hi"}],
+            model_lower="deepseek-v4-pro",
+            reasoning_config={"enabled": False},
+            max_tokens_param_fn=lambda n: {"max_tokens": n},
+        )
+        assert "reasoning_effort" not in kw
+
+    def test_deepseek_v4_thinking_enabled_extra_body(self, transport):
+        kw = transport.build_kwargs(
+            model="deepseek-v4-pro", messages=[{"role": "user", "content": "Hi"}],
+            model_lower="deepseek-v4-pro",
+            max_tokens_param_fn=lambda n: {"max_tokens": n},
+        )
+        assert kw["extra_body"]["thinking"] == {"type": "enabled"}
+
+    def test_deepseek_v4_thinking_disabled_extra_body(self, transport):
+        kw = transport.build_kwargs(
+            model="deepseek-v4-pro", messages=[{"role": "user", "content": "Hi"}],
+            model_lower="deepseek-v4-pro",
+            reasoning_config={"enabled": False},
+            max_tokens_param_fn=lambda n: {"max_tokens": n},
+        )
+        assert kw["extra_body"]["thinking"] == {"type": "disabled"}


The new DeepSeek V4 tests cover reasoning_config={"enabled": False} but not the existing disable shape reasoning_config={"effort": "none"} (used by CLI/batch runner paths). Add a regression test asserting that effort=none results in extra_body.thinking.type == "disabled" and that reasoning_effort is omitted for V4.

- _supports_reasoning_extra_body(): detect deepseek-v4 models by name - build_kwargs(): emit reasoning_effort top-level (xhigh→max, others→high) and extra_body.thinking.type=enabled/disabled for DeepSeek V4 - Exclude DeepSeek V4 from generic extra_body.reasoning path - Add 7 transport-level tests covering enabled/disabled/xhigh/chat-clean

alt-glitch · 2026-04-27T15:59:34Z

Likely duplicate of #14958 — same fix: plumb reasoning_effort and thinking toggle to DeepSeek V4 API for non-OpenRouter routes. Multiple competing PRs exist (#15577, #16448, #15251).

Copilot AI review requested due to automatic review settings April 27, 2026 15:19

Copilot started reviewing on behalf of Skyline10124 April 27, 2026 15:20 View session

Copilot AI reviewed Apr 27, 2026

View reviewed changes

Skyline10124 force-pushed the fix/deepseek-v4-thinking-mode branch from 427d9c9 to feeb834 Compare April 27, 2026 15:33

Skyline10124 requested a review from Copilot April 27, 2026 15:43

Copilot started reviewing on behalf of Skyline10124 April 27, 2026 15:43 View session

Copilot AI reviewed Apr 27, 2026

View reviewed changes

Skyline10124 force-pushed the fix/deepseek-v4-thinking-mode branch from feeb834 to 7d329ae Compare April 27, 2026 15:56

alt-glitch added type/bug Something isn't working P2 Medium — degraded but workaround exists comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint provider/deepseek DeepSeek API labels Apr 27, 2026

alt-glitch mentioned this pull request May 8, 2026

feat(opencode-go): add top-level reasoning_effort support for DeepSeek models #21577

Closed

Skyline10124 closed this May 18, 2026

Navigation Menu

Uh oh!

fix(transport): emit deepseek-v4 thinking.type and reasoning_effort on non-OpenRouter routes - #16614

fix(transport): emit deepseek-v4 thinking.type and reasoning_effort on non-OpenRouter routes#16614
Skyline10124 wants to merge 1 commit into
NousResearch:mainfrom
Skyline10124:fix/deepseek-v4-thinking-mode

Skyline10124 commented Apr 27, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 27, 2026

Uh oh!

Copilot AI Apr 27, 2026

Uh oh!

Copilot AI Apr 27, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 27, 2026

Uh oh!

Copilot AI Apr 27, 2026

Uh oh!

Copilot AI Apr 27, 2026

Uh oh!

alt-glitch commented Apr 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

	is_deepseek_v4 = params.get("model_lower", "").startswith("deepseek-v4")
	is_deepseek_v4 = model_lower.startswith("deepseek-v4")

Uh oh!

Conversation

Skyline10124 commented Apr 27, 2026

What does this PR do?

Related Issue

Type of Change

Changes Made

How to Test

Checklist

Code

Documentation & Housekeeping

Why this approach

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Copilot AI Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 27, 2026

Choose a reason for hiding this comment

Uh oh!

alt-glitch commented Apr 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants