Summary

When using DeepSeek via the Anthropic Messages-compatible endpoint (https://api.deepseek.com/anthropic) with thinking-capable models (e.g. deepseek-v4-pro), Hermes can trigger HTTP 400 from DeepSeek with an error equivalent to:

The content[].thinking in the thinking mode must be passed back to the API.

Root cause (hypothesis)

In agent/anthropic_adapter.py, convert_messages_to_anthropic() classifies api.deepseek.com as a generic third-party Anthropic endpoint (_is_third_party_anthropic_endpoint → true because the host is not anthropic.com).

For third-party hosts, the thinking-block pass strips all thinking / redacted_thinking blocks from assistant history (see the branch that builds stripped content for _is_third_party).

DeepSeek V4 thinking mode appears to require those thinking content blocks to round-trip on subsequent turns (similar in spirit to Kimi /coding, which already has a dedicated preserve path).

So: Hermes strips required blocks → next request is invalid → 400.

Expected behavior

For DeepSeek’s /anthropic route, assistant message history should preserve model-native thinking blocks across turns (at minimum the same class of handling as Kimi’s /coding endpoint: strip only Anthropic-signed blocks that third parties cannot validate, keep unsigned/native thinking).

Optionally, Hermes should also avoid stacking Anthropic-native thinking / output_config request knobs on DeepSeek’s proxy if DeepSeek drives thinking server-side (parallel to the existing Kimi /coding skip in build_anthropic_kwargs).

Actual behavior

Thinking blocks are removed from replayed assistant turns → DeepSeek rejects the follow-up request with HTTP 400.

Repro (high level)

1. Configure primary model:
   • provider: deepseek
   • model: deepseek-v4-pro (or another thinking-enabled SKU)
   • base_url: https://api.deepseek.com/anthropic
2. Enable Hermes reasoning / extended thinking path (e.g. agent.reasoning_effort: xhigh — may contribute to entering thinking mode).
3. Run a multi-turn session with tool calls (or any path that replays assistant history).
4. Observe HTTP 400 complaining that content[].thinking must be passed back.

Suggested fix direction

1. Add an endpoint detector for DeepSeek Anthropic compat (host contains api.deepseek.com and path contains /anthropic).
2. In convert_messages_to_anthropic, treat that endpoint like Kimi /coding for thinking-block preservation (do not use the "strip all thinking for all third-party messages" path).
3. In build_anthropic_kwargs, consider skipping Hermes-injected Anthropic thinking / output_config for that host (same rationale as Kimi /coding).

Workarounds (today)

• Use a non-thinking model variant if available, or
• Set agent.reasoning_effort: none (may reduce pressure but may not fully address model-driven thinking), or
• Avoid the /anthropic route if an OpenAI-compatible DeepSeek route avoids this contract.

References

• Related class of issue: DeepSeek tool-call turns needing reasoning_content padding (run_agent.py mentions DeepSeek/Kimi tool reasoning requirements).
• Kimi special-case in convert_messages_to_anthropic is a good template for DeepSeek /anthropic.

Happy to help validate a patch or add a regression test if you want a contributor to pick this up.