feat(agent): pass reasoning_effort + thinking to DeepSeek API

[zhanglpg]

Problem

DeepSeek V4 Pro/Flash thinking mode requires two parameters to control thinking effort:

1. *$reasoning_effort — top-level API parameter (high / max)
2. $thinking$ — {"type": "enabled/disabled"} in extra_body

Per DeepSeek API docs:

response = client.chat.completions.create(
    model="deepseek-v4-pro",
    reasoning_effort="max",
    extra_body={"thinking": {"type": "enabled"}}
)

Previously, Hermes only passed these parameters for Kimi (api.kimi.com / moonshot.ai). When using DeepSeek via a custom provider or the deepseek provider name, no reasoning parameters were sent — the API relied on DeepSeek's own defaults (thinking=enabled, effort=high). Users had no way to control the effort level.

Changes

3 files, +34 −2

File	Change
run_agent.py	Detect _is_deepseek (provider name deepseek or api.deepseek.com base URL) and pass to transport
agent/transports/chat_completions.py	Handle DeepSeek like Kimi: pass reasoning_effort top-level + thinking in extra_body
hermes_constants.py	Add max to VALID_REASONING_EFFORTS (DeepSeek accepts it natively)

How It Works

• User sets reasoning_effort: max in config.yaml
• Hermes detects DeepSeek provider (name deepseek or api.deepseek.com base URL)
• Sends reasoning_effort="max" + extra_body={"thinking": {"type": "enabled"}}
• DeepSeek uses max thinking effort for the response

Testing

• All 41 existing test_chat_completions.py tests pass
• 177/178 reasoning-related tests pass (1 pre-existing failure in test_deepseek_reasoning_content_echo.py, not caused by this change)

Tested on

• macOS 14 (Intel, x86_64)

[alt-glitch]

Related to #14958, #15577, #15446 — multiple competing PRs for DeepSeek V4 thinking mode support. This is the most minimal approach.

[zhanglpg]

@alt-glitch New to this repo, mind letting me know if this can be approved, or you would prefer a competing PR to be chosen. Let me know.

[teknium1]

Automated hermes-sweeper review: this DeepSeek thinking-mode support is already implemented on current main.

• plugins/model-providers/deepseek/__init__.py:47 defines DeepSeekProfile; its build_api_kwargs_extras() emits extra_body.thinking at line 67 and the top-level reasoning_effort at lines 75-80.
• agent/transports/chat_completions.py:562-594 invokes that profile hook and merges both outputs into the outgoing Chat Completions request.
• tests/plugins/model_providers/test_deepseek_profile.py:155-169 verifies the complete deepseek-v4-pro request shape, including reasoning_effort and extra_body.thinking.
• Commit cd9470f41638bd515db096cd934c463205790110 (fix(deepseek): wire thinking-mode via DeepSeekProfile, not legacy fallback) superseded the legacy transport approach and shipped in v2026.5.16.

Thank you for the focused implementation and for participating in the competing-PR discussion.