What does this PR do?

DeepSeek's direct API (api.deepseek.com) defaults to thinking=enabled when no explicit thinking parameter is sent. When the user sets reasoning_effort: none in config, the setting was silently ignored because _supports_reasoning_extra_body() returns False for direct DeepSeek endpoints — so extra_body.reasoning (OpenRouter format) was never sent. The model defaulted to thinking=enabled, generated reasoning_content in its response, and the next API call failed with HTTP 400: "reasoning_content must be passed back".

This PR adds DeepSeek direct API thinking control to ChatCompletionsTransport.build_kwargs(), following the existing Kimi pattern: emit extra_body.thinking with type: "enabled" or type: "disabled" based on the user's reasoning_config. The is_deepseek_direct flag is threaded from run_agent.py, detected via provider name (deepseek) or base URL matching (api.deepseek.com).

The OpenRouter DeepSeek path is unchanged — it uses extra_body.reasoning which is a separate parameter handled by the existing supports_reasoning gate.

Related Issue

Fixes #17212

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• agent/transports/chat_completions.py: Added DeepSeek direct API extra_body.thinking block (after existing Kimi block), with enabled/disabled control based on reasoning_config
• run_agent.py: Added _is_deepseek_direct detection (provider == "deepseek" or base_url matches api.deepseek.com) and threaded it to transport params
• tests/agent/transports/test_chat_completions.py: Added 5 test cases covering default enabled, disabled via config, effort=none, OpenRouter unchanged, Kimi unchanged

How to Test

1. Configure Hermes with DeepSeek direct API (api.deepseek.com) and reasoning_effort: none
2. Send a message that triggers tool calls (e.g. "run date")
3. Verify the second API call succeeds (no 400 error)
4. Run the targeted tests: pytest tests/agent/transports/test_chat_completions.py -v
5. Run full suite: pytest tests/ -q --ignore=tests/integration --ignore=tests/e2e

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes
• I've tested on my platform: macOS (Darwin 25.4.0, Apple Silicon), Python 3.11

Documentation & Housekeeping

• Updated relevant documentation — or N/A
• Updated cli-config.yaml.example — or N/A
• Updated contributing / agents docs — or N/A
• Considered cross-platform impact — or N/A
• Updated tool descriptions/schemas — or N/A