feat(providers): support deepseek native api reasoning_

[WuTianyi123]

What does this PR do?

Fix reasoning_effort being silently ignored when using the DeepSeek
native provider (api.deepseek.com). Two issues:

1. DeepSeek's native API uses top-level reasoning_effort (high/max),
   not extra_body.reasoning. The base ProviderProfile returns ({}, {})
   for build_api_kwargs_extras, so the config was never sent.
2. "max" was not in VALID_REASONING_EFFORTS — setting it caused
   Unknown reasoning_effort 'max', using default (medium).

Related Issue

N/A (no existing issue)

Type of Change

• ✨ New feature (non-breaking change that adds functionality)
• ✅ Tests (adding or improving test coverage)

Changes Made

• hermes_constants.py: Add "max" to VALID_REASONING_EFFORTS
• plugins/model-providers/deepseek/__init__.py: New DeepSeekProfile class
  with build_api_kwargs_extras emitting top-level reasoning_effort,
  mapping: low/medium/minimal → "high", xhigh/max → "max"
• run_agent.py: Summary path handles deepseek like LM Studio (top-level
  reasoning_effort, not extra_body.reasoning)
• cli.py: /reasoning help text updated
• cli-config.yaml.example: Options comment updated
• batch_runner.py: Docstring + valid_efforts list updated
• tests/providers/test_transport_parity.py: 7 new TestDeepSeekParity tests

How to Test

1. pytest tests/providers/test_transport_parity.py -q (26→33 tests, deepseek
   mapping coverage)
2. pytest tests/run_agent/test_run_agent.py -k reasoning -q (35 reasoning tests)
3. pytest tests/providers/ -q (90 provider tests)
4. Manual: set reasoning_effort: xhigh with provider: deepseek, verify
   reasoning_effort: max is sent in API request

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/feature
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes
• I've tested on my platform: Ubuntu 24.04

Documentation & Housekeeping

• I've updated relevant documentation (docstrings updated) — or N/A
• I've updated cli-config.yaml.example — config comment updated
• N/A — no architecture/workflow changes
• I've considered cross-platform impact — N/A (provider plugin, no OS dependency)
• N/A — no tool behavior change

Screenshots / Logs

Before (reasoning_effort: xhigh with provider: deepseek):

• _supports_reasoning_extra_body() returned False
• reasoning_effort was silently dropped from API request
  After:
• DeepSeek profile emits reasoning_effort as top-level param
• xhigh → "max", medium → "high" (per DeepSeek API spec)
• Transport parity tests verify all mappings

[teknium1]

Thanks for identifying the native DeepSeek reasoning_effort gap. The normal request path is now implemented by merged PR #26648 (cd9470f41638bd515db096cd934c463205790110): plugins/model-providers/deepseek/__init__.py:47-82 emits both top-level reasoning_effort and extra_body.thinking, and agent/transports/chat_completions.py:562-575 applies the profile extras.

Problems

• The remaining iteration-limit summary path still calls chat.completions.create() directly. It handles only LM Studio at agent/chat_completion_helpers.py:1650-1691, and repeats the same request construction on retry at :1774-1787.
• The PR's summary branch adds only reasoning_effort; it would not include DeepSeek V4's extra_body.thinking control that the canonical profile emits at plugins/model-providers/deepseek/__init__.py:67-80.

Suggested changes

• Salvage the summary-path portion into agent/chat_completion_helpers.py, preferably by deriving the summary extras from the registered provider profile so its wire shape cannot drift from the main transport.
• Add DeepSeek V4 initial-summary and retry-summary tests covering enabled and disabled reasoning.

Automated hermes-sweeper review.

[teknium1]

This bypasses the generic extra_body.reasoning path for DeepSeek, but the subsequent summary kwargs add only reasoning_effort. DeepSeek V4 also needs its extra_body.thinking control; please source both extras from the DeepSeek provider-profile contract so the direct summary request matches normal transport calls.