diff --git a/internal/agent/act.go b/internal/agent/act.go
index a1023bc8..327217d9 100644
--- a/internal/agent/act.go
+++ b/internal/agent/act.go
@@ -317,9 +317,10 @@ func (e *AgentEngine) runToolCall(
if repairErr := json.Unmarshal([]byte(repaired), &args); repairErr != nil {
logger.Errorf(ctx, "%s Failed to parse arguments (repair failed): %v", toolTag, err)
return types.ToolCall{
- ID: tc.ID,
- Name: tc.Function.Name,
- Args: map[string]any{"_raw": argsStr},
+ ID: tc.ID,
+ Name: tc.Function.Name,
+ Args: map[string]any{"_raw": argsStr},
+ ProviderMetadata: tc.ProviderMetadata,
Result: &types.ToolResult{
Success: false,
Error: fmt.Sprintf(
@@ -413,11 +414,12 @@ func (e *AgentEngine) runToolCall(
duration := time.Since(toolCallStartTime).Milliseconds()
toolCall := types.ToolCall{
- ID: tc.ID,
- Name: tc.Function.Name,
- Args: args,
- Result: result,
- Duration: duration,
+ ID: tc.ID,
+ Name: tc.Function.Name,
+ Args: args,
+ Result: result,
+ Duration: duration,
+ ProviderMetadata: tc.ProviderMetadata,
}
if err != nil {
diff --git a/internal/agent/observe.go b/internal/agent/observe.go
index e2c50bd5..363f6203 100644
--- a/internal/agent/observe.go
+++ b/internal/agent/observe.go
@@ -326,8 +326,9 @@ func (e *AgentEngine) appendToolResults(
argsJSON, _ := json.Marshal(tc.Args)
assistantMsg.ToolCalls = append(assistantMsg.ToolCalls, chat.ToolCall{
- ID: tc.ID,
- Type: "function",
+ ID: tc.ID,
+ Type: "function",
+ ProviderMetadata: tc.ProviderMetadata,
Function: chat.FunctionCall{
Name: tc.Name,
Arguments: string(argsJSON),
diff --git a/internal/agent/observe_test.go b/internal/agent/observe_test.go
index a5969241..971b9e39 100644
--- a/internal/agent/observe_test.go
+++ b/internal/agent/observe_test.go
@@ -2,6 +2,7 @@ package agent
import (
"context"
+ "encoding/json"
"testing"
"time"
@@ -73,9 +74,10 @@ func TestAppendToolResults_PreservesReasoningContent(t *testing.T) {
Thought: "I will call search.",
ReasoningContent: "Detailed chain of thought from MiMo/DeepSeek.",
ToolCalls: []types.ToolCall{{
- ID: "call_1",
- Name: "knowledge_search",
- Args: map[string]interface{}{"query": "hi"},
+ ID: "call_1",
+ Name: "knowledge_search",
+ Args: map[string]interface{}{"query": "hi"},
+ ProviderMetadata: types.ToolCallMetadata{"google": json.RawMessage(`{"thought_signature":"gemini-thought-signature"}`)},
Result: &types.ToolResult{
Success: true,
Output: "result text",
@@ -94,6 +96,8 @@ func TestAppendToolResults_PreservesReasoningContent(t *testing.T) {
"and DeepSeek thinking-mode see it on the next round (issue #1302)")
require.Len(t, out[0].ToolCalls, 1)
assert.Equal(t, "call_1", out[0].ToolCalls[0].ID)
+ assert.JSONEq(t, `{"thought_signature":"gemini-thought-signature"}`,
+ string(out[0].ToolCalls[0].ProviderMetadata["google"]))
assert.Equal(t, "tool", out[1].Role)
assert.Equal(t, "result text", out[1].Content)
diff --git a/internal/application/service/agent_history.go b/internal/application/service/agent_history.go
index 1cf0792c..7fb7f843 100644
--- a/internal/application/service/agent_history.go
+++ b/internal/application/service/agent_history.go
@@ -165,8 +165,9 @@ func buildAssistantHistoryMessages(m *types.Message) []chat.Message {
for _, tc := range nonTerminalCalls {
argsJSON, _ := json.Marshal(tc.Args)
assistantMsg.ToolCalls = append(assistantMsg.ToolCalls, chat.ToolCall{
- ID: tc.ID,
- Type: "function",
+ ID: tc.ID,
+ Type: "function",
+ ProviderMetadata: tc.ProviderMetadata,
Function: chat.FunctionCall{
Name: tc.Name,
Arguments: string(argsJSON),
diff --git a/internal/application/service/agent_history_test.go b/internal/application/service/agent_history_test.go
index 5887a81b..9b1c31dc 100644
--- a/internal/application/service/agent_history_test.go
+++ b/internal/application/service/agent_history_test.go
@@ -1,12 +1,14 @@
package service
import (
+ "encoding/json"
"testing"
agenttools "github.com/Tencent/WeKnora/internal/agent/tools"
"github.com/Tencent/WeKnora/internal/models/chat"
"github.com/Tencent/WeKnora/internal/types"
"github.com/stretchr/testify/assert"
+ "github.com/stretchr/testify/require"
)
// TestBuildUserHistoryMessage_PrefersRenderedContent verifies the user side of
@@ -247,9 +249,10 @@ func TestBuildAssistantHistoryMessages_ReplaysReasoningContent(t *testing.T) {
Thought: "Let me search.",
ReasoningContent: "model's chain of thought",
ToolCalls: []types.ToolCall{{
- ID: "call_1",
- Name: agenttools.ToolKnowledgeSearch,
- Args: map[string]interface{}{"query": "foo"},
+ ID: "call_1",
+ Name: agenttools.ToolKnowledgeSearch,
+ Args: map[string]interface{}{"query": "foo"},
+ ProviderMetadata: types.ToolCallMetadata{"google": json.RawMessage(`{"thought_signature":"gemini-history-signature"}`)},
Result: &types.ToolResult{
Success: true,
Output: "doc A",
@@ -265,6 +268,9 @@ func TestBuildAssistantHistoryMessages_ReplaysReasoningContent(t *testing.T) {
assert.Equal(t, "model's chain of thought", got[0].ReasoningContent,
"reasoning_content from AgentStep must be replayed onto the rebuilt assistant message "+
"so MiMo/DeepSeek thinking-mode does not 400 on multi-turn (issue #1302)")
+ require.Len(t, got[0].ToolCalls, 1)
+ assert.JSONEq(t, `{"thought_signature":"gemini-history-signature"}`,
+ string(got[0].ToolCalls[0].ProviderMetadata["google"]))
// Tool message and final answer message must NOT carry reasoning_content.
assert.Empty(t, got[1].ReasoningContent)
assert.Empty(t, got[2].ReasoningContent)
diff --git a/internal/models/chat/chat.go b/internal/models/chat/chat.go
index 98e7ef75..47ed6941 100644
--- a/internal/models/chat/chat.go
+++ b/internal/models/chat/chat.go
@@ -71,9 +71,10 @@ type Message struct {
// ToolCall represents a tool call in a message
type ToolCall struct {
- ID string `json:"id"`
- Type string `json:"type"` // "function"
- Function FunctionCall `json:"function"`
+ ID string `json:"id"`
+ Type string `json:"type"` // "function"
+ Function FunctionCall `json:"function"`
+ ProviderMetadata types.ToolCallMetadata `json:"provider_metadata,omitempty"`
}
// FunctionCall represents a function call
diff --git a/internal/models/chat/openai_request.go b/internal/models/chat/openai_request.go
index fff22905..85b680d5 100644
--- a/internal/models/chat/openai_request.go
+++ b/internal/models/chat/openai_request.go
@@ -1,6 +1,7 @@
package chat
import (
+ "encoding/json"
"fmt"
"github.com/sashabaranov/go-openai"
@@ -181,3 +182,61 @@ func (c *RemoteAPIChat) BuildChatCompletionRequest(
return req
}
+
+func (c *RemoteAPIChat) buildProviderOpenAIRequest(
+ body any,
+ openAIMessages []openai.ChatCompletionMessage,
+ messages []Message,
+) (map[string]any, error) {
+ data, err := json.Marshal(body)
+ if err != nil {
+ return nil, fmt.Errorf("marshal provider request: %w", err)
+ }
+
+ var out map[string]any
+ if err := json.Unmarshal(data, &out); err != nil {
+ return nil, fmt.Errorf("unmarshal provider request: %w", err)
+ }
+
+ providerMessages := make([]map[string]any, 0, len(openAIMessages))
+ for i, msg := range openAIMessages {
+ msgData, err := json.Marshal(msg)
+ if err != nil {
+ return nil, fmt.Errorf("marshal provider message: %w", err)
+ }
+ var msgMap map[string]any
+ if err := json.Unmarshal(msgData, &msgMap); err != nil {
+ return nil, fmt.Errorf("unmarshal provider message: %w", err)
+ }
+
+ if i < len(messages) && len(messages[i].ToolCalls) > 0 && len(msg.ToolCalls) > 0 {
+ toolCalls := make([]map[string]any, 0, len(msg.ToolCalls))
+ for j, tc := range msg.ToolCalls {
+ tcData, err := json.Marshal(tc)
+ if err != nil {
+ return nil, fmt.Errorf("marshal provider tool call: %w", err)
+ }
+ var tcMap map[string]any
+ if err := json.Unmarshal(tcData, &tcMap); err != nil {
+ return nil, fmt.Errorf("unmarshal provider tool call: %w", err)
+ }
+ if j < len(messages[i].ToolCalls) {
+ c.adapter.InjectToolCallMetadata(tcMap, messages[i].ToolCalls[j].ProviderMetadata)
+ }
+ toolCalls = append(toolCalls, tcMap)
+ }
+ msgMap["tool_calls"] = toolCalls
+ }
+
+ providerMessages = append(providerMessages, msgMap)
+ }
+ out["messages"] = providerMessages
+ return out, nil
+}
+
+func (c *RemoteAPIChat) shapeProviderRequest(body any, req openai.ChatCompletionRequest, messages []Message) (any, error) {
+ if !c.adapter.ForceRawHTTP() {
+ return body, nil
+ }
+ return c.buildProviderOpenAIRequest(body, req.Messages, messages)
+}
diff --git a/internal/models/chat/openai_stream.go b/internal/models/chat/openai_stream.go
index 86b1c4bb..cec9b1a5 100644
--- a/internal/models/chat/openai_stream.go
+++ b/internal/models/chat/openai_stream.go
@@ -54,6 +54,37 @@ func (c *RemoteAPIChat) parseCompletionResponse(resp *openai.ChatCompletionRespo
return response, nil
}
+func (c *RemoteAPIChat) applyCompletionToolCallMetadata(body []byte, result *types.ChatResponse) {
+ if result == nil || len(result.ToolCalls) == 0 {
+ return
+ }
+
+ var raw struct {
+ Choices []struct {
+ Message struct {
+ ToolCalls []json.RawMessage `json:"tool_calls,omitempty"`
+ } `json:"message"`
+ } `json:"choices"`
+ }
+ if err := json.Unmarshal(body, &raw); err != nil || len(raw.Choices) == 0 {
+ return
+ }
+
+ for i, rawToolCall := range raw.Choices[0].Message.ToolCalls {
+ var indexed struct {
+ Index *int `json:"index,omitempty"`
+ }
+ _ = json.Unmarshal(rawToolCall, &indexed)
+ idx := i
+ if indexed.Index != nil {
+ idx = *indexed.Index
+ }
+ if idx >= 0 && idx < len(result.ToolCalls) {
+ result.ToolCalls[idx].ProviderMetadata = c.adapter.ExtractToolCallMetadata(rawToolCall)
+ }
+ }
+}
+
// removeThinkingContent 移除思考模型输出中的 思考过程
// 仅当内容以 开头时才处理
func removeThinkingContent(content string) string {
@@ -237,11 +268,45 @@ func (c *RemoteAPIChat) processRawHTTPStream(
Delta: choice.Delta.ChatCompletionStreamChoiceDelta,
FinishReason: choice.FinishReason,
}
+ c.applyStreamToolCallMetadata(event.Data, state)
c.processStreamDelta(ctx, &sdkChoice, state, streamChan, reasoning)
}
}
}
+func (c *RemoteAPIChat) applyStreamToolCallMetadata(data []byte, state *streamState) {
+ if state == nil {
+ return
+ }
+
+ var raw struct {
+ Choices []struct {
+ Delta struct {
+ ToolCalls []json.RawMessage `json:"tool_calls,omitempty"`
+ } `json:"delta"`
+ } `json:"choices"`
+ }
+ if err := json.Unmarshal(data, &raw); err != nil || len(raw.Choices) == 0 {
+ return
+ }
+
+ for _, rawToolCall := range raw.Choices[0].Delta.ToolCalls {
+ metadata := c.adapter.ExtractToolCallMetadata(rawToolCall)
+ if len(metadata) == 0 {
+ continue
+ }
+ var indexed struct {
+ Index *int `json:"index,omitempty"`
+ }
+ _ = json.Unmarshal(rawToolCall, &indexed)
+ idx := 0
+ if indexed.Index != nil {
+ idx = *indexed.Index
+ }
+ state.setToolCallProviderMetadata(idx, metadata)
+ }
+}
+
// streamState 流式处理状态
type streamState struct {
thinkingEmitter
@@ -302,6 +367,24 @@ func (s *streamState) buildOrderedToolCalls() []types.LLMToolCall {
return result
}
+func (s *streamState) setToolCallProviderMetadata(index int, metadata types.ToolCallMetadata) {
+ if len(metadata) == 0 {
+ return
+ }
+ toolCallEntry, exists := s.toolCallMap[index]
+ if !exists || toolCallEntry == nil {
+ toolCallEntry = &types.LLMToolCall{
+ Type: "function",
+ Function: types.FunctionCall{
+ Name: "",
+ Arguments: "",
+ },
+ }
+ s.toolCallMap[index] = toolCallEntry
+ }
+ toolCallEntry.ProviderMetadata = metadata
+}
+
// processStreamDelta 处理流式响应的单个 delta
func (c *RemoteAPIChat) processStreamDelta(
ctx context.Context,
diff --git a/internal/models/chat/provider.go b/internal/models/chat/provider.go
index 4c9d3f81..c010da98 100644
--- a/internal/models/chat/provider.go
+++ b/internal/models/chat/provider.go
@@ -1,11 +1,13 @@
package chat
import (
+ "encoding/json"
"net/http"
"strings"
"github.com/Tencent/WeKnora/internal/models/provider"
modelutils "github.com/Tencent/WeKnora/internal/models/utils"
+ "github.com/Tencent/WeKnora/internal/types"
"github.com/google/uuid"
"github.com/sashabaranov/go-openai"
)
@@ -46,6 +48,12 @@ type providerAdapter interface {
// ForceRawHTTP forces the raw HTTP path even when the body is standard
// (needed by providers that must sign the exact request bytes). Default: false.
ForceRawHTTP() bool
+ // ExtractToolCallMetadata captures provider-specific state from a raw
+ // OpenAI-compatible tool_call object. Default: nil.
+ ExtractToolCallMetadata(raw json.RawMessage) types.ToolCallMetadata
+ // InjectToolCallMetadata writes provider-specific state back into an outbound
+ // OpenAI-compatible tool_call object. Default: noop.
+ InjectToolCallMetadata(toolCall map[string]any, metadata types.ToolCallMetadata)
}
// baseProvider supplies the default behavior for every providerAdapter method.
@@ -65,6 +73,10 @@ func (baseProvider) Auth(req *http.Request, creds authCreds, _ []byte) {
req.Header.Set("Authorization", "Bearer "+creds.APIKey)
}
func (baseProvider) ForceRawHTTP() bool { return false }
+func (baseProvider) ExtractToolCallMetadata(json.RawMessage) types.ToolCallMetadata {
+ return nil
+}
+func (baseProvider) InjectToolCallMetadata(map[string]any, types.ToolCallMetadata) {}
// --- WeKnoraCloud: custom endpoint + request signing + multi-content downgrade ---
@@ -152,6 +164,40 @@ type nvidiaProvider struct{ baseProvider }
func (nvidiaProvider) Name() provider.ProviderName { return provider.ProviderNvidia }
func (nvidiaProvider) Thinking() ThinkingStrategy { return chatTemplateKwargs{} }
+// --- Gemini OpenAI compatibility: tool thought signatures live in extra_content ---
+
+type geminiProvider struct{ baseProvider }
+
+func (geminiProvider) Name() provider.ProviderName { return provider.ProviderGemini }
+func (geminiProvider) ForceRawHTTP() bool { return true }
+func (geminiProvider) ExtractToolCallMetadata(raw json.RawMessage) types.ToolCallMetadata {
+ var tc struct {
+ ExtraContent map[string]json.RawMessage `json:"extra_content,omitempty"`
+ }
+ if err := json.Unmarshal(raw, &tc); err != nil {
+ return nil
+ }
+ google, ok := tc.ExtraContent["google"]
+ if !ok || len(google) == 0 {
+ return nil
+ }
+ return types.ToolCallMetadata{"google": google}
+}
+func (geminiProvider) InjectToolCallMetadata(toolCall map[string]any, metadata types.ToolCallMetadata) {
+ if len(metadata) == 0 {
+ return
+ }
+ google, ok := metadata["google"]
+ if !ok || len(google) == 0 {
+ return
+ }
+ var googleValue any
+ if err := json.Unmarshal(google, &googleValue); err != nil {
+ return
+ }
+ toolCall["extra_content"] = map[string]any{"google": googleValue}
+}
+
// --- Volcengine (火山引擎 Ark): thinking via { "thinking": { "type": ... } } ---
type volcengineProvider struct{ baseProvider }
@@ -227,6 +273,7 @@ var providerRegistry = []providerAdapter{
lkeapProvider{},
deepseekProvider{},
genericProvider{},
+ geminiProvider{},
volcengineProvider{},
nvidiaProvider{},
azureReasoningProvider{},
diff --git a/internal/models/chat/provider_test.go b/internal/models/chat/provider_test.go
index ea10dfff..6c15ce42 100644
--- a/internal/models/chat/provider_test.go
+++ b/internal/models/chat/provider_test.go
@@ -26,6 +26,7 @@ func TestResolveProvider(t *testing.T) {
{"lkeap r1 falls back", provider.ProviderLKEAP, "deepseek-r1", baseProvider{}},
{"qwen thinking", provider.ProviderAliyun, "qwen3-32b", qwenThinkingProvider{}},
{"generic", provider.ProviderGeneric, "anything", genericProvider{}},
+ {"gemini", provider.ProviderGemini, "gemini-3-flash-preview", geminiProvider{}},
{"nvidia", provider.ProviderNvidia, "anything", nvidiaProvider{}},
{"volcengine", provider.ProviderVolcengine, "doubao", volcengineProvider{}},
{"openai non-reasoning falls back", provider.ProviderOpenAI, "gpt-4o", baseProvider{}},
@@ -67,7 +68,8 @@ func TestBuildOutbound_Thinking(t *testing.T) {
t.Run("generic explicit thinking_type overrides legacy kwargs", func(t *testing.T) {
c := newOutboundChat(t, string(provider.ProviderGeneric), "deepseek-v4-flash",
map[string]string{ExtraConfigThinkingControl: "thinking_type"})
- body, _, useRaw := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(false)}, true)
+ body, _, useRaw, err := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(false)}, true)
+ require.NoError(t, err)
require.True(t, useRaw)
js := mustJSON(t, body)
assert.Contains(t, js, `"thinking"`)
@@ -77,7 +79,8 @@ func TestBuildOutbound_Thinking(t *testing.T) {
t.Run("generic legacy chat_template_kwargs", func(t *testing.T) {
c := newOutboundChat(t, string(provider.ProviderGeneric), "qwen", nil)
- body, _, useRaw := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(false)}, true)
+ body, _, useRaw, err := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(false)}, true)
+ require.NoError(t, err)
require.True(t, useRaw)
assert.Contains(t, mustJSON(t, body), "chat_template_kwargs")
})
@@ -85,7 +88,8 @@ func TestBuildOutbound_Thinking(t *testing.T) {
t.Run("none keeps the standard SDK request", func(t *testing.T) {
c := newOutboundChat(t, string(provider.ProviderGeneric), "x",
map[string]string{ExtraConfigThinkingControl: "none"})
- body, _, useRaw := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(false)}, true)
+ body, _, useRaw, err := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(false)}, true)
+ require.NoError(t, err)
assert.False(t, useRaw)
_, ok := body.(*openai.ChatCompletionRequest)
assert.True(t, ok)
@@ -93,20 +97,23 @@ func TestBuildOutbound_Thinking(t *testing.T) {
t.Run("qwen non-stream forces disabled", func(t *testing.T) {
c := newOutboundChat(t, string(provider.ProviderAliyun), "qwen3-32b", nil)
- body, _, useRaw := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(true)}, false)
+ body, _, useRaw, err := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(true)}, false)
+ require.NoError(t, err)
require.True(t, useRaw)
assert.Contains(t, mustJSON(t, body), `"enable_thinking":false`)
})
t.Run("qwen stream honors requested true", func(t *testing.T) {
c := newOutboundChat(t, string(provider.ProviderAliyun), "qwen3-32b", nil)
- body, _, _ := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(true)}, true)
+ body, _, _, err := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(true)}, true)
+ require.NoError(t, err)
assert.Contains(t, mustJSON(t, body), `"enable_thinking":true`)
})
t.Run("volcengine thinking enabled", func(t *testing.T) {
c := newOutboundChat(t, string(provider.ProviderVolcengine), "doubao", nil)
- body, _, useRaw := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(true)}, true)
+ body, _, useRaw, err := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(true)}, true)
+ require.NoError(t, err)
require.True(t, useRaw)
js := mustJSON(t, body)
assert.Contains(t, js, `"thinking"`)
@@ -115,14 +122,16 @@ func TestBuildOutbound_Thinking(t *testing.T) {
t.Run("lkeap deepseek-v3 emits thinking type", func(t *testing.T) {
c := newOutboundChat(t, string(provider.ProviderLKEAP), "deepseek-v3.1", nil)
- body, _, useRaw := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(false)}, true)
+ body, _, useRaw, err := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(false)}, true)
+ require.NoError(t, err)
require.True(t, useRaw)
assert.Contains(t, mustJSON(t, body), `"thinking"`)
})
t.Run("lkeap r1 left untouched", func(t *testing.T) {
c := newOutboundChat(t, string(provider.ProviderLKEAP), "deepseek-r1", nil)
- body, _, useRaw := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(false)}, true)
+ body, _, useRaw, err := c.buildOutbound(msgs, &ChatOptions{Thinking: ptrBool(false)}, true)
+ require.NoError(t, err)
assert.False(t, useRaw)
_, ok := body.(*openai.ChatCompletionRequest)
assert.True(t, ok)
@@ -136,7 +145,8 @@ func TestBuildOutbound_ShapeRequest(t *testing.T) {
t.Run("deepseek strips tool_choice", func(t *testing.T) {
c := newOutboundChat(t, string(provider.ProviderDeepSeek), "deepseek-chat", nil)
- body, _, useRaw := c.buildOutbound(msgs, &ChatOptions{ToolChoice: "auto"}, false)
+ body, _, useRaw, err := c.buildOutbound(msgs, &ChatOptions{ToolChoice: "auto"}, false)
+ require.NoError(t, err)
assert.False(t, useRaw)
req := body.(*openai.ChatCompletionRequest)
assert.Nil(t, req.ToolChoice)
@@ -144,13 +154,41 @@ func TestBuildOutbound_ShapeRequest(t *testing.T) {
t.Run("moonshot pins temperature to 1", func(t *testing.T) {
c := newOutboundChat(t, string(provider.ProviderMoonshot), "moonshot-v1-8k", nil)
- body, _, _ := c.buildOutbound(msgs, &ChatOptions{Temperature: 0.7, TopP: 0.9}, false)
+ body, _, _, err := c.buildOutbound(msgs, &ChatOptions{Temperature: 0.7, TopP: 0.9}, false)
+ require.NoError(t, err)
req := body.(*openai.ChatCompletionRequest)
assert.EqualValues(t, 1, req.Temperature)
assert.EqualValues(t, 0, req.TopP)
})
}
+func TestBuildOutbound_GeminiProviderMetadata(t *testing.T) {
+ c := newOutboundChat(t, string(provider.ProviderGemini), "gemini-3-flash-preview", nil)
+ messages := []Message{
+ {Role: "user", Content: "find docs"},
+ {
+ Role: "assistant",
+ ToolCalls: []ToolCall{{
+ ID: "call_1",
+ Type: "function",
+ ProviderMetadata: types.ToolCallMetadata{"google": json.RawMessage(`{"thought_signature":"gemini-signature"}`)},
+ Function: FunctionCall{
+ Name: "wiki_search",
+ Arguments: `{"query":"MACS"}`,
+ },
+ }},
+ },
+ }
+
+ body, _, useRaw, err := c.buildOutbound(messages, &ChatOptions{}, false)
+ require.NoError(t, err)
+ require.True(t, useRaw)
+
+ js := mustJSON(t, body)
+ assert.Contains(t, js, `"extra_content"`)
+ assert.Contains(t, js, `"thought_signature":"gemini-signature"`)
+}
+
func mustJSON(t *testing.T, v any) string {
t.Helper()
b, err := json.Marshal(v)
diff --git a/internal/models/chat/remote_api.go b/internal/models/chat/remote_api.go
index 7b570d32..a936406a 100644
--- a/internal/models/chat/remote_api.go
+++ b/internal/models/chat/remote_api.go
@@ -131,7 +131,7 @@ func (c *RemoteAPIChat) shapedRequest(messages []Message, opts *ChatOptions, isS
// replacing the former buildRequestCustomizer plumbing.
func (c *RemoteAPIChat) buildOutbound(
messages []Message, opts *ChatOptions, isStream bool,
-) (body any, endpoint string, useRawHTTP bool) {
+) (body any, endpoint string, useRawHTTP bool, err error) {
req := c.shapedRequest(messages, opts, isStream)
thinking := c.thinkingOverride
@@ -144,9 +144,13 @@ func (c *RemoteAPIChat) buildOutbound(
if customBody != nil {
body = customBody
}
+ body, err = c.shapeProviderRequest(body, req, messages)
+ if err != nil {
+ return nil, "", false, err
+ }
endpoint = c.adapter.Endpoint(c.baseURL, c.modelID, isStream)
useRawHTTP = useRaw || c.adapter.ForceRawHTTP() || endpoint != ""
- return body, endpoint, useRawHTTP
+ return body, endpoint, useRawHTTP, nil
}
// logRequest 记录请求日志
@@ -164,7 +168,10 @@ func (c *RemoteAPIChat) Chat(ctx context.Context, messages []Message, opts *Chat
timeoutCtx, cancel := withLLMTimeout(ctx, defaultChatTimeout)
defer cancel()
- body, endpoint, useRawHTTP := c.buildOutbound(messages, opts, false)
+ body, endpoint, useRawHTTP, err := c.buildOutbound(messages, opts, false)
+ if err != nil {
+ return nil, err
+ }
if useRawHTTP {
return c.chatWithRawHTTP(timeoutCtx, endpoint, body)
}
@@ -233,8 +240,13 @@ func (c *RemoteAPIChat) chatWithRawHTTP(ctx context.Context, endpoint string, cu
return nil, fmt.Errorf("API request failed with status %d: %s", resp.StatusCode, string(body))
}
+ body, err := io.ReadAll(resp.Body)
+ if err != nil {
+ return nil, fmt.Errorf("read response: %w", err)
+ }
+
var chatResp openai.ChatCompletionResponse
- if err := json.NewDecoder(resp.Body).Decode(&chatResp); err != nil {
+ if err := json.Unmarshal(body, &chatResp); err != nil {
return nil, fmt.Errorf("decode response: %w", err)
}
@@ -242,6 +254,7 @@ func (c *RemoteAPIChat) chatWithRawHTTP(ctx context.Context, endpoint string, cu
if err != nil {
return nil, err
}
+ c.applyCompletionToolCallMetadata(body, result)
logUsage(ctx, c.modelName, &result.Usage)
return result, nil
}
@@ -252,7 +265,11 @@ func (c *RemoteAPIChat) ChatStream(ctx context.Context, messages []Message, opts
// 因为带思考/推理的模型可能数十秒甚至几分钟才产出首 token。
timeoutCtx, cancel := withLLMTimeout(ctx, defaultStreamTimeout)
- body, endpoint, useRawHTTP := c.buildOutbound(messages, opts, true)
+ body, endpoint, useRawHTTP, err := c.buildOutbound(messages, opts, true)
+ if err != nil {
+ cancel()
+ return nil, err
+ }
if useRawHTTP {
ch, err := c.chatStreamWithRawHTTP(timeoutCtx, endpoint, body)
return wrapStreamCancel(ch, err, cancel)
diff --git a/internal/models/chat/remote_api_test.go b/internal/models/chat/remote_api_test.go
index a870f6be..db755e52 100644
--- a/internal/models/chat/remote_api_test.go
+++ b/internal/models/chat/remote_api_test.go
@@ -270,6 +270,65 @@ func TestConvertMessages_ReasoningContentRoundTrip(t *testing.T) {
})
}
+func TestApplyCompletionToolCallMetadata(t *testing.T) {
+ c := newTestRemoteChat(t)
+ c.adapter = geminiProvider{}
+
+ resp := &types.ChatResponse{
+ ToolCalls: []types.LLMToolCall{{
+ ID: "call_1",
+ Type: "function",
+ Function: types.FunctionCall{
+ Name: "wiki_search",
+ Arguments: `{"query":"MACS"}`,
+ },
+ }},
+ }
+ body := []byte(`{
+ "choices":[{
+ "message":{
+ "tool_calls":[{
+ "id":"call_1",
+ "type":"function",
+ "function":{"name":"wiki_search","arguments":"{\"query\":\"MACS\"}"},
+ "extra_content":{"google":{"thought_signature":"sig-from-gemini"}}
+ }]
+ }
+ }]
+ }`)
+
+ c.applyCompletionToolCallMetadata(body, resp)
+ require.Len(t, resp.ToolCalls, 1)
+ assert.JSONEq(t, `{"thought_signature":"sig-from-gemini"}`,
+ string(resp.ToolCalls[0].ProviderMetadata["google"]))
+}
+
+func TestApplyStreamToolCallMetadata(t *testing.T) {
+ c := newTestRemoteChat(t)
+ c.adapter = geminiProvider{}
+ state := newStreamState()
+
+ body := []byte(`{
+ "choices":[{
+ "delta":{
+ "tool_calls":[{
+ "index":0,
+ "id":"call_1",
+ "type":"function",
+ "function":{"name":"wiki_search","arguments":"{\"query\":\"MACS\"}"},
+ "extra_content":{"google":{"thought_signature":"stream-sig-from-gemini"}}
+ }]
+ }
+ }]
+ }`)
+
+ c.applyStreamToolCallMetadata(body, state)
+ toolCalls := state.buildOrderedToolCalls()
+ require.Len(t, toolCalls, 1)
+ assert.JSONEq(t, `{"thought_signature":"stream-sig-from-gemini"}`,
+ string(toolCalls[0].ProviderMetadata["google"]))
+}
+
// TestRemoteAPIChat 综合测试 Remote API Chat 的所有功能
func TestRemoteAPIChat(t *testing.T) {
// 获取环境变量
diff --git a/internal/types/agent.go b/internal/types/agent.go
index b5f2d1a7..e5e3cb87 100644
--- a/internal/types/agent.go
+++ b/internal/types/agent.go
@@ -176,12 +176,13 @@ type ToolResult struct {
// ToolCall represents a single tool invocation within an agent step
type ToolCall struct {
- ID string `json:"id"` // Function call ID from LLM
- Name string `json:"name"` // Tool name
- Args map[string]interface{} `json:"args"` // Tool arguments
- Result *ToolResult `json:"result"` // Execution result (contains Output)
- Reflection string `json:"reflection,omitempty"` // Agent's reflection on this tool call result (if enabled)
- Duration int64 `json:"duration"` // Execution time in milliseconds
+ ID string `json:"id"` // Function call ID from LLM
+ Name string `json:"name"` // Tool name
+ Args map[string]interface{} `json:"args"` // Tool arguments
+ Result *ToolResult `json:"result"` // Execution result (contains Output)
+ Reflection string `json:"reflection,omitempty"` // Agent's reflection on this tool call result (if enabled)
+ Duration int64 `json:"duration"` // Execution time in milliseconds
+ ProviderMetadata ToolCallMetadata `json:"provider_metadata,omitempty"` // Provider-specific tool-call state for replay
}
// AgentStep represents one iteration of the ReAct loop
diff --git a/internal/types/chat.go b/internal/types/chat.go
index 5839d8d9..d4b0a2fa 100644
--- a/internal/types/chat.go
+++ b/internal/types/chat.go
@@ -34,11 +34,16 @@ type TokenUsage struct {
// LLMToolCall represents a function/tool call from the LLM
type LLMToolCall struct {
- ID string `json:"id"`
- Type string `json:"type"` // "function"
- Function FunctionCall `json:"function"`
+ ID string `json:"id"`
+ Type string `json:"type"` // "function"
+ Function FunctionCall `json:"function"`
+ ProviderMetadata ToolCallMetadata `json:"provider_metadata,omitempty"`
}
+// ToolCallMetadata carries provider-specific tool-call state that must round-trip
+// with the assistant tool call, without teaching core agent code vendor fields.
+type ToolCallMetadata map[string]json.RawMessage
+
// FunctionCall represents the function details
type FunctionCall struct {
Name string `json:"name"`