DeepSeek V4 Thinking 400 Fix for Hermes

Problem

When DeepSeek V4 Flash/Pro runs with thinking mode enabled, agentic tool calls can fail with 400 after multiple turns. This is not only a Hermes issue; Cursor, Continue.dev, RooCode, LangChain, and OpenCode can hit the same class of bug. The likely cause is that DeepSeek adds a private reasoning_content field on top of the OpenAI-compatible protocol, and many frameworks do not handle that field consistently during message round-trips, tool calls, persistence, or model switching.

How It Works

The skill normalizes outgoing messages before they reach the DeepSeek API:
- Backfills reasoning_content: assistant messages always carry the field, preserving existing content or using an empty value.
- Deep-copies state: reduces accidental mutation when multiple functions share the same dictionary.
- Cleans vendor fields: removes reasoning_content when switching to non-thinking models to prevent leakage.
- Repairs persisted fields: keeps both reasoning and reasoning_content when saving, then restores missing fields on load.
- Sends explicit thinking control: when thinking is off, it passes a disabled value instead of relying on the provider default.

Boundaries

This is aimed at Hermes Agent plus DeepSeek V4 Flash/Pro in multi-turn tool-call loops, cron auxiliary paths, and restored sessions. It does not replace model reasoning, fix network or quota problems, or guarantee compatibility with future API changes; re-validation is needed if the protocol evolves.