fix: ensure reasoning_content consistency for DeepSeek-compatible APIs

[highland0971]

Summary

DeepSeek V4 thinking mode (and compatible APIs like Ark Coding Plan at ark.cn-beijing.volces.com) requires reasoning_content on every assistant message in the conversation history. An inconsistent mix — some messages with reasoning_content, some without — triggers HTTP 400:

The reasoning_content in the thinking mode must be passed back to the API.

This bug manifests on the Ark Coding Plan endpoint when models intermittently return reasoning_content (e.g., glm-5.1), producing 22 messages with it and 71 without in the same conversation.

Changes

environments/agent_loop.py (subagent loop / HermesAgentLoop)

• Always set reasoning_content on assistant messages (use empty string when no reasoning was returned in that turn)
• Previously only set it when reasoning was non-None, creating inconsistent messages

run_agent.py — _build_assistant_message()

• Extend defensive padding to cover all assistant messages for DeepSeek/Kimi, not just tool-call ones
• Previously: elif msg.get("tool_calls") and self._needs_deepseek_tool_reasoning()
• Now: elif self._needs_deepseek_tool_reasoning() or self._needs_kimi_tool_reasoning()

run_agent.py — _copy_reasoning_content_for_api() (replay path)

• Same change: pad all assistant messages during replay, not just tool-call ones

run_agent.py — _needs_deepseek_tool_reasoning() (detection)

• Add detection for Ark Coding Plan endpoint (ark.cn-beijing.volces.com) which follows DeepSeek API conventions

Tests

• Update test_deepseek_assistant_no_tool_call_now_padded to expect padding on non-tool-call messages
• Add test_ark_coding_plan_endpoint for Ark detection

Test Plan

All 22 tests in test_deepseek_reasoning_content_echo.py pass. All 7 reasoning-related tests in test_hermes_state.py pass.

Refs #15250

[alt-glitch]

Related to #15446 (comprehensive DeepSeek V4 support) and refs #15250. This PR extends reasoning_content padding to non-tool-call assistant messages and adds Ark Coding Plan endpoint detection.

[teknium1]

Closing as redundant — the DeepSeek reasoning_content thinking-mode 400 and cross-provider leak chain of issues is now fully covered on main:

• Direct provider detection + same-provider tool-call pad: commits 93a2d6b, d58b305, ad0ac89, 5ae6081
• All-assistant-messages pad rule: commit ad0ac89
• Ordering / cross-provider isolation: commit 9daa062 + regression guard 63bf7a2
• Cross-provider leak (MiniMax reasoning → DeepSeek): PR fix(agent): block cross-provider reasoning leak to DeepSeek/Kimi (#15748) #16500 (merging shortly)

21 regression tests in tests/run_agent/test_deepseek_reasoning_content_echo.py + 2 new tests for the cross-provider scenario exercise every known path. Thanks for the submission — appreciate the digging on this area.