DeepSeek tool-calling assistant turns missing reasoning_content cause HTTP 400 — Kimi-only compensation should generalise

Summary

run_agent.py::_copy_reasoning_content_for_api includes a special case that injects an empty reasoning_content string on prior assistant turns that contain tool_calls but no reasoning field, but only when the active provider is Kimi/Moonshot. DeepSeek's reasoning-capable models (deepseek-reasoner, deepseek-chat with thinking on) require the same compensation — without it the next turn is rejected with HTTP 400 ("messages with tool_calls must include reasoning_content").

Tested on Hermes Agent v0.11.0.

Reproduction

config.yaml:

model:
  default: deepseek-reasoner
  provider: deepseek

Send a message that triggers a tool call. On the second turn (when the model would normally see its previous tool-call assistant message in the replay) the API returns:

HTTP 400 - messages with tool_calls must include reasoning_content

Switch to kimi-coding with the same conversation shape and the call succeeds — because _copy_reasoning_content_for_api injects the missing field for Kimi.

Root Cause

run_agent.py::_copy_reasoning_content_for_api (around the version shipped in v0.11.0):

def _copy_reasoning_content_for_api(self, source_msg: dict, api_msg: dict) -> None:
    """Copy provider-facing reasoning fields onto an API replay message."""
    if source_msg.get("role") != "assistant":
        return

    explicit_reasoning = source_msg.get("reasoning_content")
    if isinstance(explicit_reasoning, str):
        api_msg["reasoning_content"] = explicit_reasoning
        return

    normalized_reasoning = source_msg.get("reasoning")
    if isinstance(normalized_reasoning, str) and normalized_reasoning:
        api_msg["reasoning_content"] = normalized_reasoning
        return

    kimi_requires_reasoning = (
        self.provider in {"kimi-coding", "kimi-coding-cn"}
        or base_url_host_matches(self.base_url, "api.kimi.com")
        or base_url_host_matches(self.base_url, "moonshot.ai")
        or base_url_host_matches(self.base_url, "moonshot.cn")
    )
    if kimi_requires_reasoning and source_msg.get("tool_calls"):
        api_msg["reasoning_content"] = ""

The compensation block only fires for Kimi/Moonshot endpoints. DeepSeek imposes the same constraint when reasoning mode is active but is not in the allowlist, so the missing-reasoning_content assistant turn slips through and the upstream rejects it.

Suggested Fix

Generalise the compensation to cover DeepSeek when reasoning is enabled:

deepseek_requires_reasoning = (
    self.provider == "deepseek"
    or base_url_host_matches(self.base_url, "api.deepseek.com")
)

if (kimi_requires_reasoning or deepseek_requires_reasoning) and source_msg.get("tool_calls"):
    api_msg["reasoning_content"] = ""

Or, more cleanly, move the allowlist to a module-level constant and add deepseek (plus api.deepseek.com) to it.

Impact

DeepSeek is a popular, low-cost provider for users who want native reasoning + tool use. Without this fix, any DeepSeek-backed Hermes agent that uses tools breaks on the second turn — effectively unusable for anything beyond a single-turn chat.

Workaround

Switch reasoning off (use deepseek-chat without thinking), or route DeepSeek through OpenRouter / a relay that strips the reasoning_content requirement upstream. Both lose the value DeepSeek offers in the first place.

Related

The same pattern likely applies to other reasoning-capable third-party endpoints that mirror the OpenAI Chat Completions schema (Z.AI / GLM, MiniMax M2 thinking, Qwen3 with thinking on, etc.). Each would need a similar allowlist entry — or, better, a unified rule keyed on reasoning_enabled && previous_turn_has_tool_calls && previous_turn_lacks_reasoning_content.