fix(agent): pass reasoning_content back to providers that require it

MiMo and DeepSeek V3.2/V4 reject multi-turn requests in thinking mode
when the prior assistant message lacks reasoning_content with HTTP 400:
"The reasoning_content in the thinking mode must be passed back to the API."

Agent's ReAct loop is the worst-case scenario — every round produces tool
calls, exactly the case DeepSeek's docs specify reasoning_content MUST
participate in subsequent context.

Plumb reasoning_content through the full assistant-message round-trip:
- chat.Message / types.ChatResponse / types.AgentStep gain a
  reasoning_content field (AgentStep persists via the existing
  Message.AgentSteps jsonb column, no migration needed).
- streamLLMToEventBus accumulates reasoning chunks into
  result.ReasoningContent and surfaces it on the round's ChatResponse.
- engine.runReActIteration writes it onto AgentStep so cross-turn replay
  preserves it.
- observe.appendToolResults attaches it to the same-turn assistant
  message; agent_history.buildAssistantHistoryMessages does the same on
  cross-turn replay.
- RemoteAPIChat.ConvertMessages forwards it on assistant turns to
  openai.ChatCompletionMessage.ReasoningContent (already supported by
  go-openai); providers that don't recognize the field ignore it.

Tests cover the three boundaries: ConvertMessages serializes it for
upstream, appendToolResults preserves it within the same turn, and
buildAssistantHistoryMessages replays it across turns.

Scope is intentionally limited to Agent mode — KnowledgeQA's chat
pipeline and Anthropic's signed thinking_blocks are separate fixes that
require schema changes (rendered_content / thinking_blocks columns).

Fixes #1302
This commit is contained in:
wizardchen
2026-05-13 12:26:11 +08:00
committed by lyingbug
parent 1ae06fb857
commit b00bc84f35
11 changed files with 225 additions and 28 deletions
+9 -4
View File
@@ -186,10 +186,15 @@ type ToolCall struct {
// AgentStep represents one iteration of the ReAct loop
type AgentStep struct {
Iteration int `json:"iteration"` // Iteration number (0-indexed)
Thought string `json:"thought"` // LLM's reasoning/thinking (Think phase)
ToolCalls []ToolCall `json:"tool_calls"` // Tools called in this step (Act phase)
Timestamp time.Time `json:"timestamp"` // When this step occurred
Iteration int `json:"iteration"` // Iteration number (0-indexed)
Thought string `json:"thought"` // LLM's reasoning/thinking (Think phase)
// ReasoningContent stores the OpenAI-protocol reasoning_content emitted by the
// model in this round. Persisted on AgentStep so cross-turn replay can put it
// back on the assistant message — required by MiMo / DeepSeek V3.2+ thinking
// mode, ignored by providers that don't recognize the field.
ReasoningContent string `json:"reasoning_content,omitempty"`
ToolCalls []ToolCall `json:"tool_calls"` // Tools called in this step (Act phase)
Timestamp time.Time `json:"timestamp"` // When this step occurred
}
// GetObservations returns observations from all tool calls in this step
+7 -4
View File
@@ -27,10 +27,13 @@ type FunctionCall struct {
// ChatResponse chat response
type ChatResponse struct {
Content string `json:"content"`
ToolCalls []LLMToolCall `json:"tool_calls,omitempty"`
FinishReason string `json:"finish_reason,omitempty"`
Usage TokenUsage `json:"usage"`
Content string `json:"content"`
// ReasoningContent 是支持思考链的模型(DeepSeek thinking、小米 MiMo、vLLM reasoning 等)
// 在本轮输出的推理内容。需要在后续多轮请求中原样回传给那些严格校验的供应商。
ReasoningContent string `json:"reasoning_content,omitempty"`
ToolCalls []LLMToolCall `json:"tool_calls,omitempty"`
FinishReason string `json:"finish_reason,omitempty"`
Usage TokenUsage `json:"usage"`
}
// Response type