Skip to content

Database governance built for humans and agents — controlling changes and access across every major database.

feat(review): fail a review whose reply was truncated or blocked (#21591)

h3n4l1b251a510eb5
diff --git a/backend/api/mcp/gen/openapi.yaml b/backend/api/mcp/gen/openapi.yaml
index a8053288e8..5b73c78429 100644
--- a/backend/api/mcp/gen/openapi.yaml
+++ b/backend/api/mcp/gen/openapi.yaml
@@ -9079,9 +9079,25 @@ components:
- type: "null"
title: usage
description: Token usage for this provider call, when available.
+ stopReason:
+ allOf:
+ - $ref: '#/components/schemas/bytebase.v1.AIChatResponse.StopReason'
+ title: stop_reason
+ description: Why the model stopped. A truncated or blocked response can still carry content.
title: AIChatResponse
additionalProperties: false
description: Response message for AIService.Chat.
+ bytebase.v1.AIChatResponse.StopReason:
+ type: string
+ title: StopReason
+ enum:
+ - STOP_REASON_UNSPECIFIED
+ - TURN_COMPLETED
+ - TOKEN_LIMIT
+ - BLOCKED
+ - MALFORMED_TOOL_CALL
+ - OTHER
+ description: Why the model stopped generating a response.
bytebase.v1.AIChatToolCall:
type: object
properties:
diff --git a/backend/generated-go/v1/ai_service.pb.go b/backend/generated-go/v1/ai_service.pb.go
index 50492bd87a..32ff09df5f 100644
--- a/backend/generated-go/v1/ai_service.pb.go
+++ b/backend/generated-go/v1/ai_service.pb.go
@@ -83,6 +83,71 @@ func (AIChatMessageRole) EnumDescriptor() ([]byte, []int) {
return file_v1_ai_service_proto_rawDescGZIP(), []int{0}
}
+// Why the model stopped generating a response.
+type AIChatResponse_StopReason int32
+
+const (
+ // The provider did not report why the model stopped.
+ AIChatResponse_STOP_REASON_UNSPECIFIED AIChatResponse_StopReason = 0
+ // The model completed its turn with an answer or with tool calls.
+ AIChatResponse_TURN_COMPLETED AIChatResponse_StopReason = 1
+ // The response was truncated at a token limit and is incomplete.
+ AIChatResponse_TOKEN_LIMIT AIChatResponse_StopReason = 2
+ // The provider blocked the request or the response, or the model refused to answer.
+ AIChatResponse_BLOCKED AIChatResponse_StopReason = 3
+ // The model generated a tool call that the provider could not parse.
+ AIChatResponse_MALFORMED_TOOL_CALL AIChatResponse_StopReason = 4
+ // The model stopped for a reason that no other value names.
+ AIChatResponse_OTHER AIChatResponse_StopReason = 5
+)
+
+// Enum value maps for AIChatResponse_StopReason.
+var (
+ AIChatResponse_StopReason_name = map[int32]string{
+ 0: "STOP_REASON_UNSPECIFIED",
+ 1: "TURN_COMPLETED",
+ 2: "TOKEN_LIMIT",
+ 3: "BLOCKED",
+ 4: "MALFORMED_TOOL_CALL",
+ 5: "OTHER",
+ }
+ AIChatResponse_StopReason_value = map[string]int32{
+ "STOP_REASON_UNSPECIFIED": 0,
+ "TURN_COMPLETED": 1,
+ "TOKEN_LIMIT": 2,
+ "BLOCKED": 3,
+ "MALFORMED_TOOL_CALL": 4,
+ "OTHER": 5,
+ }
+)
+
+func (x AIChatResponse_StopReason) Enum() *AIChatResponse_StopReason {
+ p := new(AIChatResponse_StopReason)
+ *p = x
+ return p
+}
+
+func (x AIChatResponse_StopReason) String() string {
+ return protoimpl.X.EnumStringOf(x.Descriptor(), protoreflect.EnumNumber(x))
+}
+
+func (AIChatResponse_StopReason) Descriptor() protoreflect.EnumDescriptor {
+ return file_v1_ai_service_proto_enumTypes[1].Descriptor()
+}
+
+func (AIChatResponse_StopReason) Type() protoreflect.EnumType {
+ return &file_v1_ai_service_proto_enumTypes[1]
+}
+
+func (x AIChatResponse_StopReason) Number() protoreflect.EnumNumber {
+ return protoreflect.EnumNumber(x)
+}
+
+// Deprecated: Use AIChatResponse_StopReason.Descriptor instead.
+func (AIChatResponse_StopReason) EnumDescriptor() ([]byte, []int) {
+ return file_v1_ai_service_proto_rawDescGZIP(), []int{5, 0}
+}
+
// A single message in the conversation.
type AIChatMessage struct {
state protoimpl.MessageState `protogen:"open.v1"`
@@ -403,7 +468,9 @@ type AIChatResponse struct {
// Tool calls the AI wants to make.
ToolCalls []*AIChatToolCall `protobuf:"bytes,2,rep,name=tool_calls,json=toolCalls,proto3" json:"tool_calls,omitempty"`
// Token usage for this provider call, when available.
- Usage *AIChatUsage `protobuf:"bytes,3,opt,name=usage,proto3,oneof" json:"usage,omitempty"`
+ Usage *AIChatUsage `protobuf:"bytes,3,opt,name=usage,proto3,oneof" json:"usage,omitempty"`
+ // Why the model stopped. A truncated or blocked response can still carry content.
+ StopReason AIChatResponse_StopReason `protobuf:"varint,4,opt,name=stop_reason,json=stopReason,proto3,enum=bytebase.v1.AIChatResponse_StopReason" json:"stop_reason,omitempty"`
unknownFields protoimpl.UnknownFields
sizeCache protoimpl.SizeCache
}
@@ -459,6 +526,13 @@ func (x *AIChatResponse) GetUsage() *AIChatUsage {
return nil
}
+func (x *AIChatResponse) GetStopReason() AIChatResponse_StopReason {
+ if x != nil {
+ return x.StopReason
+ }
+ return AIChatResponse_STOP_REASON_UNSPECIFIED
+}
+
var File_v1_ai_service_proto protoreflect.FileDescriptor
const file_v1_ai_service_proto_rawDesc = "" +
@@ -488,12 +562,22 @@ const file_v1_ai_service_proto_rawDesc = "" +
"\bmessages\x18\x01 \x03(\v2\x1a.bytebase.v1.AIChatMessageB\x04\xd0\xea0\x02R\bmessages\x12R\n" +
"\x10tool_definitions\x18\x02 \x03(\v2!.bytebase.v1.AIChatToolDefinitionB\x04\xd0\xea0\x02R\x0ftoolDefinitions\"0\n" +
"\vAIChatUsage\x12!\n" +
- "\ftotal_tokens\x18\x01 \x01(\x05R\vtotalTokens\"\xb6\x01\n" +
+ "\ftotal_tokens\x18\x01 \x01(\x05R\vtotalTokens\"\x80\x03\n" +
"\x0eAIChatResponse\x12\x1d\n" +
"\acontent\x18\x01 \x01(\tH\x00R\acontent\x88\x01\x01\x12:\n" +
"\n" +
"tool_calls\x18\x02 \x03(\v2\x1b.bytebase.v1.AIChatToolCallR\ttoolCalls\x123\n" +
- "\x05usage\x18\x03 \x01(\v2\x18.bytebase.v1.AIChatUsageH\x01R\x05usage\x88\x01\x01B\n" +
+ "\x05usage\x18\x03 \x01(\v2\x18.bytebase.v1.AIChatUsageH\x01R\x05usage\x88\x01\x01\x12G\n" +
+ "\vstop_reason\x18\x04 \x01(\x0e2&.bytebase.v1.AIChatResponse.StopReasonR\n" +
+ "stopReason\"\x7f\n" +
+ "\n" +
+ "StopReason\x12\x1b\n" +
+ "\x17STOP_REASON_UNSPECIFIED\x10\x00\x12\x12\n" +
+ "\x0eTURN_COMPLETED\x10\x01\x12\x0f\n" +
+ "\vTOKEN_LIMIT\x10\x02\x12\v\n" +
+ "\aBLOCKED\x10\x03\x12\x17\n" +
+ "\x13MALFORMED_TOOL_CALL\x10\x04\x12\t\n" +
+ "\x05OTHER\x10\x05B\n" +
"\n" +
"\b_contentB\b\n" +
"\x06_usage*\xbc\x01\n" +
@@ -519,31 +603,33 @@ func file_v1_ai_service_proto_rawDescGZIP() []byte {
return file_v1_ai_service_proto_rawDescData
}
-var file_v1_ai_service_proto_enumTypes = make([]protoimpl.EnumInfo, 1)
+var file_v1_ai_service_proto_enumTypes = make([]protoimpl.EnumInfo, 2)
var file_v1_ai_service_proto_msgTypes = make([]protoimpl.MessageInfo, 6)
var file_v1_ai_service_proto_goTypes = []any{
- (AIChatMessageRole)(0), // 0: bytebase.v1.AIChatMessageRole
- (*AIChatMessage)(nil), // 1: bytebase.v1.AIChatMessage
- (*AIChatToolDefinition)(nil), // 2: bytebase.v1.AIChatToolDefinition
- (*AIChatToolCall)(nil), // 3: bytebase.v1.AIChatToolCall
- (*AIChatRequest)(nil), // 4: bytebase.v1.AIChatRequest
- (*AIChatUsage)(nil), // 5: bytebase.v1.AIChatUsage
- (*AIChatResponse)(nil), // 6: bytebase.v1.AIChatResponse
+ (AIChatMessageRole)(0), // 0: bytebase.v1.AIChatMessageRole
+ (AIChatResponse_StopReason)(0), // 1: bytebase.v1.AIChatResponse.StopReason
+ (*AIChatMessage)(nil), // 2: bytebase.v1.AIChatMessage
+ (*AIChatToolDefinition)(nil), // 3: bytebase.v1.AIChatToolDefinition
+ (*AIChatToolCall)(nil), // 4: bytebase.v1.AIChatToolCall
+ (*AIChatRequest)(nil), // 5: bytebase.v1.AIChatRequest
+ (*AIChatUsage)(nil), // 6: bytebase.v1.AIChatUsage
+ (*AIChatResponse)(nil), // 7: bytebase.v1.AIChatResponse
}
var file_v1_ai_service_proto_depIdxs = []int32{
0, // 0: bytebase.v1.AIChatMessage.role:type_name -> bytebase.v1.AIChatMessageRole
- 3, // 1: bytebase.v1.AIChatMessage.tool_calls:type_name -> bytebase.v1.AIChatToolCall
- 1, // 2: bytebase.v1.AIChatRequest.messages:type_name -> bytebase.v1.AIChatMessage
- 2, // 3: bytebase.v1.AIChatRequest.tool_definitions:type_name -> bytebase.v1.AIChatToolDefinition
- 3, // 4: bytebase.v1.AIChatResponse.tool_calls:type_name -> bytebase.v1.AIChatToolCall
- 5, // 5: bytebase.v1.AIChatResponse.usage:type_name -> bytebase.v1.AIChatUsage
- 4, // 6: bytebase.v1.AIService.Chat:input_type -> bytebase.v1.AIChatRequest
- 6, // 7: bytebase.v1.AIService.Chat:output_type -> bytebase.v1.AIChatResponse
- 7, // [7:8] is the sub-list for method output_type
- 6, // [6:7] is the sub-list for method input_type
- 6, // [6:6] is the sub-list for extension type_name
- 6, // [6:6] is the sub-list for extension extendee
- 0, // [0:6] is the sub-list for field type_name
+ 4, // 1: bytebase.v1.AIChatMessage.tool_calls:type_name -> bytebase.v1.AIChatToolCall
+ 2, // 2: bytebase.v1.AIChatRequest.messages:type_name -> bytebase.v1.AIChatMessage
+ 3, // 3: bytebase.v1.AIChatRequest.tool_definitions:type_name -> bytebase.v1.AIChatToolDefinition
+ 4, // 4: bytebase.v1.AIChatResponse.tool_calls:type_name -> bytebase.v1.AIChatToolCall
+ 6, // 5: bytebase.v1.AIChatResponse.usage:type_name -> bytebase.v1.AIChatUsage
+ 1, // 6: bytebase.v1.AIChatResponse.stop_reason:type_name -> bytebase.v1.AIChatResponse.StopReason
+ 5, // 7: bytebase.v1.AIService.Chat:input_type -> bytebase.v1.AIChatRequest
+ 7, // 8: bytebase.v1.AIService.Chat:output_type -> bytebase.v1.AIChatResponse
+ 8, // [8:9] is the sub-list for method output_type
+ 7, // [7:8] is the sub-list for method input_type
+ 7, // [7:7] is the sub-list for extension type_name
+ 7, // [7:7] is the sub-list for extension extendee
+ 0, // [0:7] is the sub-list for field type_name
}
func init() { file_v1_ai_service_proto_init() }
@@ -560,7 +646,7 @@ func file_v1_ai_service_proto_init() {
File: protoimpl.DescBuilder{
GoPackagePath: reflect.TypeOf(x{}).PkgPath(),
RawDescriptor: unsafe.Slice(unsafe.StringData(file_v1_ai_service_proto_rawDesc), len(file_v1_ai_service_proto_rawDesc)),
- NumEnums: 1,
+ NumEnums: 2,
NumMessages: 6,
NumExtensions: 0,
NumServices: 1,
diff --git a/backend/generated-go/v1/ai_service_equal.pb.go b/backend/generated-go/v1/ai_service_equal.pb.go
index d341601f49..0f2e458b2a 100644
--- a/backend/generated-go/v1/ai_service_equal.pb.go
+++ b/backend/generated-go/v1/ai_service_equal.pb.go
@@ -131,5 +131,8 @@ func (x *AIChatResponse) Equal(y *AIChatResponse) bool {
if !x.Usage.Equal(y.Usage) {
return false
}
+ if x.StopReason != y.StopReason {
+ return false
+ }
return true
}
diff --git a/backend/plugin/ai/claude.go b/backend/plugin/ai/claude.go
index 79583d5e4b..2ac748d5fe 100644
--- a/backend/plugin/ai/claude.go
+++ b/backend/plugin/ai/claude.go
@@ -51,8 +51,9 @@ type chatClaudeUsage struct {
}
type chatClaudeResponse struct {
- Content []json.RawMessage `json:"content"`
- Usage *chatClaudeUsage `json:"usage,omitempty"`
+ Content []json.RawMessage `json:"content"`
+ StopReason string `json:"stop_reason"`
+ Usage *chatClaudeUsage `json:"usage,omitempty"`
}
func claudeContentFromToolCalls(toolCalls []*v1pb.AIChatToolCall) []json.RawMessage {
@@ -244,7 +245,7 @@ func chatClaude(ctx context.Context, aiSetting *storepb.AISetting, request *v1pb
return nil, errors.Errorf("failed to unmarshal Claude response: %s", err)
}
- result := &v1pb.AIChatResponse{}
+ result := &v1pb.AIChatResponse{StopReason: claudeStopReason(resp.StopReason)}
if resp.Usage != nil {
result.Usage = newAIChatUsage(resp.Usage.InputTokens + resp.Usage.OutputTokens)
}
diff --git a/backend/plugin/ai/gemini.go b/backend/plugin/ai/gemini.go
index 2a1098401e..b96f611211 100644
--- a/backend/plugin/ai/gemini.go
+++ b/backend/plugin/ai/gemini.go
@@ -70,7 +70,11 @@ type chatGeminiResponse struct {
Content struct {
Parts []chatGeminiResponsePart `json:"parts"`
} `json:"content"`
+ FinishReason string `json:"finishReason"`
} `json:"candidates"`
+ PromptFeedback struct {
+ BlockReason string `json:"blockReason"`
+ } `json:"promptFeedback"`
UsageMetadata *chatGeminiUsage `json:"usageMetadata,omitempty"`
}
@@ -128,6 +132,17 @@ func readGeminiFailure(httpErr *HTTPError) {
}
}
+// appendGeminiContent adds parts as a turn of role. Gemini expects the roles
+// to alternate, so parts with the role of the last turn join that turn, as the
+// results of parallel function calls do in Google's examples.
+func appendGeminiContent(contents []chatGeminiContent, role string, parts ...chatGeminiPart) []chatGeminiContent {
+ if last := len(contents) - 1; last >= 0 && contents[last].Role == role {
+ contents[last].Parts = append(contents[last].Parts, parts...)
+ return contents
+ }
+ return append(contents, chatGeminiContent{Role: role, Parts: parts})
+}
+
func chatGemini(ctx context.Context, aiSetting *storepb.AISetting, request *v1pb.AIChatRequest) (*v1pb.AIChatResponse, error) {
payload := chatGeminiRequest{}
@@ -139,10 +154,7 @@ func chatGemini(ctx context.Context, aiSetting *storepb.AISetting, request *v1pb
systemParts = append(systemParts, chatGeminiPart{Text: m.GetContent()})
}
case v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_USER:
- payload.Contents = append(payload.Contents, chatGeminiContent{
- Role: "user",
- Parts: []chatGeminiPart{{Text: m.GetContent()}},
- })
+ payload.Contents = appendGeminiContent(payload.Contents, "user", chatGeminiPart{Text: m.GetContent()})
case v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_ASSISTANT:
var parts []chatGeminiPart
if m.Content != nil && *m.Content != "" {
@@ -164,24 +176,16 @@ func chatGemini(ctx context.Context, aiSetting *storepb.AISetting, request *v1pb
}
parts = append(parts, part)
}
- payload.Contents = append(payload.Contents, chatGeminiContent{
- Role: "model",
- Parts: parts,
- })
+ payload.Contents = appendGeminiContent(payload.Contents, "model", parts...)
case v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_TOOL:
// Gemini uses functionResponse parts in a "user" role message.
// We need the tool name; get it from the tool call ID by searching prior messages.
toolName := getToolNameFromMessages(request.Messages, m.GetToolCallId())
resultData := geminiFunctionResponse(m.GetContent())
- payload.Contents = append(payload.Contents, chatGeminiContent{
- Role: "user",
- Parts: []chatGeminiPart{
- {
- FunctionResponse: &chatGeminiFunctionResult{
- Name: toolName,
- Response: resultData,
- },
- },
+ payload.Contents = appendGeminiContent(payload.Contents, "user", chatGeminiPart{
+ FunctionResponse: &chatGeminiFunctionResult{
+ Name: toolName,
+ Response: resultData,
},
})
default:
@@ -243,6 +247,7 @@ func chatGemini(ctx context.Context, aiSetting *storepb.AISetting, request *v1pb
result.Usage = newAIChatUsage(resp.UsageMetadata.TotalTokenCount)
}
if len(resp.Candidates) > 0 {
+ result.StopReason = geminiStopReason(resp.Candidates[0].FinishReason)
var textContent string
for _, part := range resp.Candidates[0].Content.Parts {
if part.Text != "" {
@@ -271,6 +276,8 @@ func chatGemini(ctx context.Context, aiSetting *storepb.AISetting, request *v1pb
if textContent != "" {
result.Content = &textContent
}
+ } else {
+ result.StopReason = geminiPromptBlockStopReason(resp.PromptFeedback.BlockReason)
}
return result, nil
}
diff --git a/backend/plugin/ai/gemini_contents_test.go b/backend/plugin/ai/gemini_contents_test.go
new file mode 100644
index 0000000000..843a9928e5
--- /dev/null
+++ b/backend/plugin/ai/gemini_contents_test.go
@@ -0,0 +1,99 @@
+package ai
+
+import (
+ "context"
+ "encoding/json"
+ "net/http"
+ "net/http/httptest"
+ "testing"
+
+ "github.com/stretchr/testify/require"
+
+ storepb "github.com/bytebase/bytebase/backend/generated-go/store"
+ v1pb "github.com/bytebase/bytebase/backend/generated-go/v1"
+)
+
+func sentGeminiRequest(t *testing.T, messages []*v1pb.AIChatMessage) chatGeminiRequest {
+ t.Helper()
+
+ var payload chatGeminiRequest
+ server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+ require.NoError(t, json.NewDecoder(r.Body).Decode(&payload))
+ w.Header().Set("Content-Type", "application/json")
+ _, err := w.Write([]byte(`{"candidates": [{"content": {"parts": [{"text": "done"}]}, "finishReason": "STOP"}]}`))
+ require.NoError(t, err)
+ }))
+ defer server.Close()
+
+ _, err := Chat(
+ context.Background(),
+ &storepb.AISetting{Provider: storepb.AISetting_GEMINI, Endpoint: server.URL, Model: "gemini-3.5-flash", ApiKey: "test-key"},
+ &v1pb.AIChatRequest{Messages: messages},
+ )
+ require.NoError(t, err)
+ return payload
+}
+
+func geminiRoles(contents []chatGeminiContent) []string {
+ roles := make([]string, 0, len(contents))
+ for _, content := range contents {
+ roles = append(roles, content.Role)
+ }
+ return roles
+}
+
+func TestChatGeminiJoinsToolResultsAndTheTextAfterThemInOneUserTurn(t *testing.T) {
+ t.Parallel()
+
+ system, prompt, correction := "You review SQL.", "Review this change.", "Call the tool again."
+ readID, searchID := "call-read", "call-search"
+ readResult, searchResult := `{"rows": 7}`, `{"total": 2}`
+ payload := sentGeminiRequest(t, []*v1pb.AIChatMessage{
+ {Role: v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_SYSTEM, Content: &system},
+ {Role: v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_USER, Content: &prompt},
+ {Role: v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_ASSISTANT, ToolCalls: []*v1pb.AIChatToolCall{
+ {Id: readID, Name: "read", Arguments: `{"objects": [{"name": "orders"}]}`},
+ {Id: searchID, Name: "search", Arguments: `{"text": "orders"}`},
+ }},
+ {Role: v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_TOOL, Content: &readResult, ToolCallId: &readID},
+ {Role: v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_TOOL, Content: &searchResult, ToolCallId: &searchID},
+ {Role: v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_USER, Content: &correction},
+ })
+
+ require.NotNil(t, payload.SystemInstruction)
+ require.Len(t, payload.SystemInstruction.Parts, 1)
+ require.Equal(t, system, payload.SystemInstruction.Parts[0].Text)
+
+ require.Equal(t, []string{"user", "model", "user"}, geminiRoles(payload.Contents))
+ require.Len(t, payload.Contents[0].Parts, 1)
+ require.Equal(t, prompt, payload.Contents[0].Parts[0].Text)
+ calls := payload.Contents[1].Parts
+ require.Len(t, calls, 2)
+ require.Equal(t, "read", calls[0].FunctionCall.Name)
+ require.Equal(t, "search", calls[1].FunctionCall.Name)
+
+ results := payload.Contents[2].Parts
+ require.Len(t, results, 3)
+ require.Equal(t, "read", results[0].FunctionResponse.Name)
+ require.Equal(t, map[string]any{"rows": float64(7)}, results[0].FunctionResponse.Response)
+ require.Equal(t, "search", results[1].FunctionResponse.Name)
+ require.Equal(t, map[string]any{"total": float64(2)}, results[1].FunctionResponse.Response)
+ require.Nil(t, results[2].FunctionResponse)
+ require.Equal(t, correction, results[2].Text)
+}
+
+func TestChatGeminiJoinsATextAfterThePromptInOneUserTurn(t *testing.T) {
+ t.Parallel()
+
+ prompt, correction := "Review this change.", "Call the tool again."
+ payload := sentGeminiRequest(t, []*v1pb.AIChatMessage{
+ {Role: v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_USER, Content: &prompt},
+ {Role: v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_USER, Content: &correction},
+ })
+
+ require.Nil(t, payload.SystemInstruction)
+ require.Equal(t, []string{"user"}, geminiRoles(payload.Contents))
+ require.Len(t, payload.Contents[0].Parts, 2)
+ require.Equal(t, prompt, payload.Contents[0].Parts[0].Text)
+ require.Equal(t, correction, payload.Contents[0].Parts[1].Text)
+}
diff --git a/backend/plugin/ai/openai.go b/backend/plugin/ai/openai.go
index e08c45c419..2402c782d6 100644
--- a/backend/plugin/ai/openai.go
+++ b/backend/plugin/ai/openai.go
@@ -62,8 +62,10 @@ type chatOpenAIResponse struct {
Message struct {
Role string `json:"role"`
Content *string `json:"content"`
+ Refusal string `json:"refusal"`
ToolCalls []json.RawMessage `json:"tool_calls"`
} `json:"message"`
+ FinishReason string `json:"finish_reason"`
} `json:"choices"`
Usage *chatOpenAIUsage `json:"usage,omitempty"`
}
@@ -318,6 +320,7 @@ func chatOpenAIChatCompletions(ctx context.Context, aiSetting *storepb.AISetting
}
if len(resp.Choices) > 0 {
msg := resp.Choices[0].Message
+ result.StopReason = openAIChatCompletionsStopReason(resp.Choices[0].FinishReason, msg.Refusal)
result.Content = msg.Content
for _, rawTC := range msg.ToolCalls {
var parsed chatOpenAIToolCallParsed
@@ -442,7 +445,7 @@ func chatOpenAIResponses(ctx context.Context, aiSetting *storepb.AISetting, requ
return nil, errors.Errorf("OpenAI Responses API returned an incomplete response: %s", reason)
}
- result := &v1pb.AIChatResponse{}
+ result := &v1pb.AIChatResponse{StopReason: openAIResponsesStopReason(response.Status)}
if response.Usage != nil {
result.Usage = newAIChatUsage(response.Usage.TotalTokens)
}
@@ -460,6 +463,7 @@ func chatOpenAIResponses(ctx context.Context, aiSetting *storepb.AISetting, requ
_, _ = content.WriteString(part.Text)
case "refusal":
_, _ = content.WriteString(part.Refusal)
+ result.StopReason = v1pb.AIChatResponse_BLOCKED
default:
continue
}
diff --git a/backend/plugin/ai/stop_reason.go b/backend/plugin/ai/stop_reason.go
new file mode 100644
index 0000000000..06ad9a8c4f
--- /dev/null
+++ b/backend/plugin/ai/stop_reason.go
@@ -0,0 +1,96 @@
+package ai
+
+import (
+ v1pb "github.com/bytebase/bytebase/backend/generated-go/v1"
+)
+
+// Each mapping sends a value its provider's API reference does not list to
+// OTHER, never to TURN_COMPLETED: a caller trusts a completed turn, and nobody
+// has read what an unlisted value means.
+
+// openAIChatCompletionsStopReason maps choices[].finish_reason. A refusal
+// arrives as message.refusal beside a finish_reason of stop.
+func openAIChatCompletionsStopReason(finishReason string, refusal string) v1pb.AIChatResponse_StopReason {
+ if refusal != "" {
+ return v1pb.AIChatResponse_BLOCKED
+ }
+ switch finishReason {
+ case "":
+ return v1pb.AIChatResponse_STOP_REASON_UNSPECIFIED
+ case "stop", "tool_calls", "function_call":
+ return v1pb.AIChatResponse_TURN_COMPLETED
+ case "length":
+ return v1pb.AIChatResponse_TOKEN_LIMIT
+ case "content_filter":
+ return v1pb.AIChatResponse_BLOCKED
+ default:
+ return v1pb.AIChatResponse_OTHER
+ }
+}
+
+// openAIResponsesStopReason maps status. The caller returns an error for an
+// incomplete response and checks for a refusal, which arrives as a content part.
+func openAIResponsesStopReason(status string) v1pb.AIChatResponse_StopReason {
+ switch status {
+ case "":
+ return v1pb.AIChatResponse_STOP_REASON_UNSPECIFIED
+ case "completed":
+ return v1pb.AIChatResponse_TURN_COMPLETED
+ default:
+ // failed, cancelled, queued, and in_progress land here.
+ return v1pb.AIChatResponse_OTHER
+ }
+}
+
+// claudeStopReason maps stop_reason.
+func claudeStopReason(stopReason string) v1pb.AIChatResponse_StopReason {
+ switch stopReason {
+ case "":
+ return v1pb.AIChatResponse_STOP_REASON_UNSPECIFIED
+ case "end_turn", "stop_sequence", "tool_use":
+ return v1pb.AIChatResponse_TURN_COMPLETED
+ case "max_tokens", "model_context_window_exceeded":
+ return v1pb.AIChatResponse_TOKEN_LIMIT
+ case "refusal":
+ return v1pb.AIChatResponse_BLOCKED
+ default:
+ // pause_turn lands here.
+ return v1pb.AIChatResponse_OTHER
+ }
+}
+
+// geminiStopReason maps candidates[].finishReason.
+func geminiStopReason(finishReason string) v1pb.AIChatResponse_StopReason {
+ switch finishReason {
+ case "", "FINISH_REASON_UNSPECIFIED":
+ return v1pb.AIChatResponse_STOP_REASON_UNSPECIFIED
+ case "STOP":
+ return v1pb.AIChatResponse_TURN_COMPLETED
+ case "MAX_TOKENS", "CONTINUATION":
+ return v1pb.AIChatResponse_TOKEN_LIMIT
+ case "SAFETY", "RECITATION", "LANGUAGE", "BLOCKLIST", "PROHIBITED_CONTENT", "SPII",
+ "IMAGE_SAFETY", "IMAGE_PROHIBITED_CONTENT", "IMAGE_RECITATION",
+ "ESCALATION", "PUP_LIMITED_DISABLED":
+ return v1pb.AIChatResponse_BLOCKED
+ case "MALFORMED_FUNCTION_CALL":
+ return v1pb.AIChatResponse_MALFORMED_TOOL_CALL
+ default:
+ // OTHER, IMAGE_OTHER, NO_IMAGE, UNEXPECTED_TOOL_CALL, TOO_MANY_TOOL_CALLS,
+ // MISSING_THOUGHT_SIGNATURE, and MALFORMED_RESPONSE land here.
+ return v1pb.AIChatResponse_OTHER
+ }
+}
+
+// geminiPromptBlockStopReason maps promptFeedback.blockReason, which Gemini
+// sets when it blocked the prompt and returned no candidate. Its OTHER is a
+// block for a reason Gemini does not name.
+func geminiPromptBlockStopReason(blockReason string) v1pb.AIChatResponse_StopReason {
+ switch blockReason {
+ case "", "BLOCK_REASON_UNSPECIFIED":
+ return v1pb.AIChatResponse_STOP_REASON_UNSPECIFIED
+ case "SAFETY", "OTHER", "BLOCKLIST", "PROHIBITED_CONTENT", "IMAGE_SAFETY":
+ return v1pb.AIChatResponse_BLOCKED
+ default:
+ return v1pb.AIChatResponse_OTHER
+ }
+}
diff --git a/backend/plugin/ai/stop_reason_test.go b/backend/plugin/ai/stop_reason_test.go
new file mode 100644
index 0000000000..a0dfedcc48
--- /dev/null
+++ b/backend/plugin/ai/stop_reason_test.go
@@ -0,0 +1,493 @@
+package ai
+
+import (
+ "context"
+ "fmt"
+ "net/http"
+ "net/http/httptest"
+ "testing"
+
+ "github.com/stretchr/testify/require"
+
+ storepb "github.com/bytebase/bytebase/backend/generated-go/store"
+ v1pb "github.com/bytebase/bytebase/backend/generated-go/v1"
+)
+
+const (
+ stopUnspecified = v1pb.AIChatResponse_STOP_REASON_UNSPECIFIED
+ stopTurnCompleted = v1pb.AIChatResponse_TURN_COMPLETED
+ stopTokenLimit = v1pb.AIChatResponse_TOKEN_LIMIT
+ stopBlocked = v1pb.AIChatResponse_BLOCKED
+ stopMalformedToolCall = v1pb.AIChatResponse_MALFORMED_TOOL_CALL
+ stopOther = v1pb.AIChatResponse_OTHER
+)
+
+// chatWithScriptedReply sends one user message through Chat to a server that
+// answers every request with reply. endpointPath selects the OpenAI wire format.
+func chatWithScriptedReply(t *testing.T, provider storepb.AISetting_Provider, endpointPath string, reply string) (*v1pb.AIChatResponse, error) {
+ t.Helper()
+
+ server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+ w.Header().Set("Content-Type", "application/json")
+ _, err := w.Write([]byte(reply))
+ require.NoError(t, err)
+ }))
+ t.Cleanup(server.Close)
+
+ content := "Review this change."
+ return Chat(
+ context.Background(),
+ &storepb.AISetting{Provider: provider, Endpoint: server.URL + endpointPath, Model: "test-model", ApiKey: "test-key"},
+ &v1pb.AIChatRequest{Messages: []*v1pb.AIChatMessage{{Role: v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_USER, Content: &content}}},
+ )
+}
+
+func TestOpenAIChatCompletionsStopReason(t *testing.T) {
+ t.Parallel()
+
+ tests := []struct {
+ name string
+ provider storepb.AISetting_Provider
+ reply string
+ want v1pb.AIChatResponse_StopReason
+ wantContent string
+ }{
+ {
+ name: "stop",
+ reply: `{"choices": [{"message": {"role": "assistant", "content": "done"}, "finish_reason": "stop"}]}`,
+ want: stopTurnCompleted,
+ },
+ {
+ name: "tool_calls",
+ reply: `{"choices": [{"message": {"role": "assistant", "content": null, "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{}"}}]}, "finish_reason": "tool_calls"}]}`,
+ want: stopTurnCompleted,
+ },
+ {
+ name: "function_call",
+ reply: `{"choices": [{"message": {"role": "assistant", "content": null}, "finish_reason": "function_call"}]}`,
+ want: stopTurnCompleted,
+ },
+ {
+ name: "length keeps the truncated text",
+ reply: `{"choices": [{"message": {"role": "assistant", "content": "{\"findings\": ["}, "finish_reason": "length"}]}`,
+ want: stopTokenLimit,
+ wantContent: `{"findings": [`,
+ },
+ {
+ name: "length with tool calls",
+ reply: `{"choices": [{"message": {"role": "assistant", "content": null, "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "read", "arguments": "{\"objects\": ["}}]}, "finish_reason": "length"}]}`,
+ want: stopTokenLimit,
+ },
+ {
+ name: "content_filter",
+ reply: `{"choices": [{"message": {"role": "assistant", "content": null}, "finish_reason": "content_filter"}]}`,
+ want: stopBlocked,
+ },
+ {
+ name: "content_filter from Azure OpenAI",
+ provider: storepb.AISetting_AZURE_OPENAI,
+ reply: `{"choices": [{"message": {"role": "assistant", "content": ""}, "finish_reason": "content_filter"}]}`,
+ want: stopBlocked,
+ },
+ {
+ name: "refusal beside a stop",
+ reply: `{"choices": [{"message": {"role": "assistant", "content": null, "refusal": "I can't help with that request."}, "finish_reason": "stop"}]}`,
+ want: stopBlocked,
+ },
+ {
+ name: "null refusal is not a refusal",
+ reply: `{"choices": [{"message": {"role": "assistant", "content": "done", "refusal": null}, "finish_reason": "stop"}]}`,
+ want: stopTurnCompleted,
+ },
+ {
+ name: "empty refusal is not a refusal",
+ reply: `{"choices": [{"message": {"role": "assistant", "content": "done", "refusal": ""}, "finish_reason": "stop"}]}`,
+ want: stopTurnCompleted,
+ },
+ {
+ name: "null finish_reason",
+ reply: `{"choices": [{"message": {"role": "assistant", "content": "done"}, "finish_reason": null}]}`,
+ want: stopUnspecified,
+ },
+ {
+ name: "no finish_reason",
+ reply: `{"choices": [{"message": {"role": "assistant", "content": "done"}}]}`,
+ want: stopUnspecified,
+ },
+ {
+ name: "no choices",
+ reply: `{"choices": []}`,
+ want: stopUnspecified,
+ },
+ {
+ name: "unknown finish_reason",
+ reply: `{"choices": [{"message": {"role": "assistant", "content": "done"}, "finish_reason": "insufficient_system_resource"}]}`,
+ want: stopOther,
+ },
+ }
+ for _, tc := range tests {
+ t.Run(tc.name, func(t *testing.T) {
+ t.Parallel()
+
+ provider := tc.provider
+ if provider == storepb.AISetting_PROVIDER_UNSPECIFIED {
+ provider = storepb.AISetting_OPEN_AI
+ }
+ response, err := chatWithScriptedReply(t, provider, "/chat/completions", tc.reply)
+ require.NoError(t, err)
+ require.Equal(t, tc.want, response.GetStopReason())
+ if tc.wantContent != "" {
+ require.Equal(t, tc.wantContent, response.GetContent())
+ }
+ })
+ }
+}
+
+func TestOpenAIResponsesStopReason(t *testing.T) {
+ t.Parallel()
+
+ const textOutput = `"output": [{"type": "message", "role": "assistant", "content": [{"type": "output_text", "text": "{\"findings\": ["}]}]`
+ const refusalOutput = `"output": [{"type": "message", "role": "assistant", "content": [{"type": "refusal", "refusal": "I can't help with that request."}]}]`
+ tests := []struct {
+ name string
+ reply string
+ want v1pb.AIChatResponse_StopReason
+ wantContent string
+ wantErrText string
+ }{
+ {
+ name: "completed",
+ reply: `{"status": "completed", ` + textOutput + `}`,
+ want: stopTurnCompleted,
+ },
+ {
+ name: "completed with a function call",
+ reply: `{"status": "completed", "output": [{"type": "function_call", "call_id": "call_1", "name": "read", "arguments": "{}"}]}`,
+ want: stopTurnCompleted,
+ },
+ {
+ name: "no status",
+ reply: `{` + textOutput + `}`,
+ want: stopUnspecified,
+ },
+ {
+ name: "refusal part of a completed response keeps the refusal text",
+ reply: `{"status": "completed", ` + refusalOutput + `}`,
+ want: stopBlocked,
+ wantContent: "I can't help with that request.",
+ },
+ {
+ name: "refusal part without a status",
+ reply: `{` + refusalOutput + `}`,
+ want: stopBlocked,
+ },
+ {
+ name: "failed without an error object",
+ reply: `{"status": "failed", ` + textOutput + `}`,
+ want: stopOther,
+ },
+ {
+ name: "cancelled",
+ reply: `{"status": "cancelled", ` + textOutput + `}`,
+ want: stopOther,
+ },
+ {
+ name: "queued",
+ reply: `{"status": "queued", "output": []}`,
+ want: stopOther,
+ },
+ {
+ name: "in_progress",
+ reply: `{"status": "in_progress", "output": []}`,
+ want: stopOther,
+ },
+ {
+ name: "unknown status",
+ reply: `{"status": "expired", ` + textOutput + `}`,
+ want: stopOther,
+ },
+ {
+ name: "incomplete at max_output_tokens is an error",
+ reply: `{"status": "incomplete", "incomplete_details": {"reason": "max_output_tokens"}, ` + textOutput + `}`,
+ wantErrText: "OpenAI Responses API returned an incomplete response: max_output_tokens",
+ },
+ {
+ name: "incomplete for content_filter is an error",
+ reply: `{"status": "incomplete", "incomplete_details": {"reason": "content_filter"}, "output": []}`,
+ wantErrText: "OpenAI Responses API returned an incomplete response: content_filter",
+ },
+ {
+ name: "incomplete with a refusal part is an error",
+ reply: `{"status": "incomplete", "incomplete_details": {"reason": "content_filter"}, ` + refusalOutput + `}`,
+ wantErrText: "OpenAI Responses API returned an incomplete response: content_filter",
+ },
+ {
+ name: "incomplete at max_messages is an error",
+ reply: `{"status": "incomplete", "incomplete_details": {"reason": "max_messages"}, ` + textOutput + `}`,
+ wantErrText: "OpenAI Responses API returned an incomplete response: max_messages",
+ },
+ {
+ name: "incomplete when steered is an error",
+ reply: `{"status": "incomplete", "incomplete_details": {"reason": "steered"}, ` + textOutput + `}`,
+ wantErrText: "OpenAI Responses API returned an incomplete response: steered",
+ },
+ {
+ name: "incomplete without details is an error",
+ reply: `{"status": "incomplete", ` + textOutput + `}`,
+ wantErrText: "OpenAI Responses API returned an incomplete response: ",
+ },
+ }
+ for _, tc := range tests {
+ t.Run(tc.name, func(t *testing.T) {
+ t.Parallel()
+
+ response, err := chatWithScriptedReply(t, storepb.AISetting_OPEN_AI, "/responses", tc.reply)
+ if tc.wantErrText != "" {
+ require.ErrorContains(t, err, tc.wantErrText)
+ require.Nil(t, response)
+ return
+ }
+ require.NoError(t, err)
+ require.Equal(t, tc.want, response.GetStopReason())
+ if tc.wantContent != "" {
+ require.Equal(t, tc.wantContent, response.GetContent())
+ }
+ })
+ }
+}
+
+func TestClaudeStopReason(t *testing.T) {
+ t.Parallel()
+
+ tests := []struct {
+ name string
+ reply string
+ want v1pb.AIChatResponse_StopReason
+ wantContent string
+ }{
+ {
+ name: "end_turn",
+ reply: `{"content": [{"type": "text", "text": "done"}], "stop_reason": "end_turn"}`,
+ want: stopTurnCompleted,
+ },
+ {
+ name: "stop_sequence",
+ reply: `{"content": [{"type": "text", "text": "done"}], "stop_reason": "stop_sequence"}`,
+ want: stopTurnCompleted,
+ },
+ {
+ name: "tool_use",
+ reply: `{"content": [{"type": "tool_use", "id": "toolu_1", "name": "read", "input": {}}], "stop_reason": "tool_use"}`,
+ want: stopTurnCompleted,
+ },
+ {
+ name: "max_tokens keeps the truncated text",
+ reply: `{"content": [{"type": "text", "text": "{\"findings\": ["}], "stop_reason": "max_tokens"}`,
+ want: stopTokenLimit,
+ wantContent: `{"findings": [`,
+ },
+ {
+ name: "max_tokens with a tool call",
+ reply: `{"content": [{"type": "tool_use", "id": "toolu_1", "name": "read", "input": {}}], "stop_reason": "max_tokens"}`,
+ want: stopTokenLimit,
+ },
+ {
+ name: "model_context_window_exceeded",
+ reply: `{"content": [{"type": "text", "text": "{\"findings\": ["}], "stop_reason": "model_context_window_exceeded"}`,
+ want: stopTokenLimit,
+ },
+ {
+ name: "refusal keeps the text before it",
+ reply: `{"content": [{"type": "text", "text": "I can't help with that request."}], "stop_reason": "refusal", "stop_details": {"type": "refusal", "category": "cyber", "explanation": null}}`,
+ want: stopBlocked,
+ wantContent: "I can't help with that request.",
+ },
+ {
+ name: "refusal without content",
+ reply: `{"content": [], "stop_reason": "refusal"}`,
+ want: stopBlocked,
+ },
+ {
+ name: "pause_turn",
+ reply: `{"content": [{"type": "text", "text": "done"}], "stop_reason": "pause_turn"}`,
+ want: stopOther,
+ },
+ {
+ name: "null stop_reason",
+ reply: `{"content": [{"type": "text", "text": "done"}], "stop_reason": null}`,
+ want: stopUnspecified,
+ },
+ {
+ name: "no stop_reason",
+ reply: `{"content": [{"type": "text", "text": "done"}]}`,
+ want: stopUnspecified,
+ },
+ {
+ name: "unknown stop_reason",
+ reply: `{"content": [{"type": "text", "text": "done"}], "stop_reason": "compaction"}`,
+ want: stopOther,
+ },
+ }
+ for _, tc := range tests {
+ t.Run(tc.name, func(t *testing.T) {
+ t.Parallel()
+
+ response, err := chatWithScriptedReply(t, storepb.AISetting_CLAUDE, "", tc.reply)
+ require.NoError(t, err)
+ require.Equal(t, tc.want, response.GetStopReason())
+ if tc.wantContent != "" {
+ require.Equal(t, tc.wantContent, response.GetContent())
+ }
+ })
+ }
+}
+
+func TestGeminiStopReasonForEveryFinishReason(t *testing.T) {
+ t.Parallel()
+
+ tests := []struct {
+ finishReason string
+ want v1pb.AIChatResponse_StopReason
+ }{
+ {finishReason: "FINISH_REASON_UNSPECIFIED", want: stopUnspecified},
+ {finishReason: "STOP", want: stopTurnCompleted},
+ {finishReason: "MAX_TOKENS", want: stopTokenLimit},
+ {finishReason: "CONTINUATION", want: stopTokenLimit},
+ {finishReason: "SAFETY", want: stopBlocked},
+ {finishReason: "RECITATION", want: stopBlocked},
+ {finishReason: "LANGUAGE", want: stopBlocked},
+ {finishReason: "BLOCKLIST", want: stopBlocked},
+ {finishReason: "PROHIBITED_CONTENT", want: stopBlocked},
+ {finishReason: "SPII", want: stopBlocked},
+ {finishReason: "IMAGE_SAFETY", want: stopBlocked},
+ {finishReason: "IMAGE_PROHIBITED_CONTENT", want: stopBlocked},
+ {finishReason: "IMAGE_RECITATION", want: stopBlocked},
+ {finishReason: "ESCALATION", want: stopBlocked},
+ {finishReason: "PUP_LIMITED_DISABLED", want: stopBlocked},
+ {finishReason: "MALFORMED_FUNCTION_CALL", want: stopMalformedToolCall},
+ {finishReason: "OTHER", want: stopOther},
+ {finishReason: "IMAGE_OTHER", want: stopOther},
+ {finishReason: "NO_IMAGE", want: stopOther},
+ {finishReason: "UNEXPECTED_TOOL_CALL", want: stopOther},
+ {finishReason: "TOO_MANY_TOOL_CALLS", want: stopOther},
+ {finishReason: "MISSING_THOUGHT_SIGNATURE", want: stopOther},
+ {finishReason: "MALFORMED_RESPONSE", want: stopOther},
+ {finishReason: "A_REASON_ADDED_LATER", want: stopOther},
+ }
+ for _, tc := range tests {
+ t.Run(tc.finishReason, func(t *testing.T) {
+ t.Parallel()
+
+ reply := fmt.Sprintf(`{"candidates": [{"content": {"parts": [{"text": "{\"findings\": ["}]}, "finishReason": %q}]}`, tc.finishReason)
+ response, err := chatWithScriptedReply(t, storepb.AISetting_GEMINI, "", reply)
+ require.NoError(t, err)
+ require.Equal(t, tc.want, response.GetStopReason())
+ require.Equal(t, `{"findings": [`, response.GetContent())
+ })
+ }
+}
+
+func TestGeminiStopReasonForCandidateShapes(t *testing.T) {
+ t.Parallel()
+
+ tests := []struct {
+ name string
+ reply string
+ want v1pb.AIChatResponse_StopReason
+ }{
+ {
+ name: "no finishReason",
+ reply: `{"candidates": [{"content": {"parts": [{"text": "done"}]}}]}`,
+ want: stopUnspecified,
+ },
+ {
+ name: "function call that stops with STOP",
+ reply: `{"candidates": [{"content": {"parts": [{"functionCall": {"name": "read", "args": {}}}]}, "finishReason": "STOP"}]}`,
+ want: stopTurnCompleted,
+ },
+ {
+ name: "safety block without content",
+ reply: `{"candidates": [{"finishReason": "SAFETY", "safetyRatings": [{"category": "HARM_CATEGORY_HATE_SPEECH", "probability": "HIGH"}]}]}`,
+ want: stopBlocked,
+ },
+ {
+ name: "malformed function call without content",
+ reply: `{"candidates": [{"content": {}, "finishReason": "MALFORMED_FUNCTION_CALL", "finishMessage": "Malformed function call: read(objects="}]}`,
+ want: stopMalformedToolCall,
+ },
+ }
+ for _, tc := range tests {
+ t.Run(tc.name, func(t *testing.T) {
+ t.Parallel()
+
+ response, err := chatWithScriptedReply(t, storepb.AISetting_GEMINI, "", tc.reply)
+ require.NoError(t, err)
+ require.Equal(t, tc.want, response.GetStopReason())
+ })
+ }
+}
+
+func TestGeminiStopReasonWithoutACandidate(t *testing.T) {
+ t.Parallel()
+
+ tests := []struct {
+ name string
+ reply string
+ want v1pb.AIChatResponse_StopReason
+ }{
+ {
+ name: "no prompt feedback",
+ reply: `{"usageMetadata": {"totalTokenCount": 12}}`,
+ want: stopUnspecified,
+ },
+ {
+ name: "prompt feedback without a block reason",
+ reply: `{"promptFeedback": {"safetyRatings": []}}`,
+ want: stopUnspecified,
+ },
+ {
+ name: "prompt blocked with BLOCK_REASON_UNSPECIFIED",
+ reply: `{"promptFeedback": {"blockReason": "BLOCK_REASON_UNSPECIFIED"}}`,
+ want: stopUnspecified,
+ },
+ {
+ name: "prompt blocked for SAFETY",
+ reply: `{"promptFeedback": {"blockReason": "SAFETY"}}`,
+ want: stopBlocked,
+ },
+ {
+ name: "prompt blocked for OTHER",
+ reply: `{"promptFeedback": {"blockReason": "OTHER"}}`,
+ want: stopBlocked,
+ },
+ {
+ name: "prompt blocked for BLOCKLIST",
+ reply: `{"promptFeedback": {"blockReason": "BLOCKLIST"}}`,
+ want: stopBlocked,
+ },
+ {
+ name: "prompt blocked for PROHIBITED_CONTENT",
+ reply: `{"promptFeedback": {"blockReason": "PROHIBITED_CONTENT"}}`,
+ want: stopBlocked,
+ },
+ {
+ name: "prompt blocked for IMAGE_SAFETY",
+ reply: `{"promptFeedback": {"blockReason": "IMAGE_SAFETY"}}`,
+ want: stopBlocked,
+ },
+ {
+ name: "unknown block reason",
+ reply: `{"promptFeedback": {"blockReason": "A_REASON_ADDED_LATER"}}`,
+ want: stopOther,
+ },
+ }
+ for _, tc := range tests {
+ t.Run(tc.name, func(t *testing.T) {
+ t.Parallel()
+
+ response, err := chatWithScriptedReply(t, storepb.AISetting_GEMINI, "", tc.reply)
+ require.NoError(t, err)
+ require.Equal(t, tc.want, response.GetStopReason())
+ })
+ }
+}
diff --git a/backend/plugin/aireview/model_ai_gemini_test.go b/backend/plugin/aireview/model_ai_gemini_test.go
new file mode 100644
index 0000000000..fa67d7ca09
--- /dev/null
+++ b/backend/plugin/aireview/model_ai_gemini_test.go
@@ -0,0 +1,50 @@
+package aireview
+
+import (
+ "context"
+ "encoding/json"
+ "net/http"
+ "net/http/httptest"
+ "testing"
+
+ "github.com/stretchr/testify/require"
+
+ storepb "github.com/bytebase/bytebase/backend/generated-go/store"
+)
+
+func TestSettingModelKeepsGeminiRolesAlternatingAfterAnUnparsedToolCall(t *testing.T) {
+ t.Parallel()
+
+ replies := []string{
+ `{"candidates": [{"content": {"parts": [{"functionCall": {"name": "search", "args": {"text": "orders"}}}]}, "finishReason": "STOP"}]}`,
+ `{"candidates": [{"content": {}, "finishReason": "MALFORMED_FUNCTION_CALL", "finishMessage": "Malformed function call: read(objects="}]}`,
+ `{"candidates": [{"content": {"parts": [{"text": "{\"findings\": []}"}]}, "finishReason": "STOP"}]}`,
+ }
+ var requests []geminiRequest
+ server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+ var request geminiRequest
+ require.NoError(t, json.NewDecoder(r.Body).Decode(&request))
+ requests = append(requests, request)
+ w.Header().Set("Content-Type", "application/json")
+ _, err := w.Write([]byte(replies[min(len(requests), len(replies))-1]))
+ require.NoError(t, err)
+ }))
+ defer server.Close()
+
+ model := NewModel(&storepb.AISetting{Provider: storepb.AISetting_GEMINI, Endpoint: server.URL, Model: "gemini-3.5-flash", ApiKey: "test-key"})
+ result, err := NewReviewer(model).Review(context.Background(), &Request{Statement: "DROP TABLE orders;"}, ordersCatalog())
+ require.NoError(t, err)
+ require.Empty(t, result.Findings)
+ require.Len(t, requests, 3)
+
+ contents := requests[2].Contents
+ roles := make([]string, 0, len(contents))
+ for _, content := range contents {
+ roles = append(roles, content.Role)
+ }
+ require.Equal(t, []string{"user", "model", "user"}, roles)
+ last := contents[2].Parts
+ require.Len(t, last, 2)
+ require.Equal(t, "search", last[0].FunctionResponse.Name)
+ require.Equal(t, malformedToolCallMessage, last[1].Text)
+}
diff --git a/backend/plugin/aireview/model_ai_test.go b/backend/plugin/aireview/model_ai_test.go
index 3f3014db33..51ed5785c3 100644
--- a/backend/plugin/aireview/model_ai_test.go
+++ b/backend/plugin/aireview/model_ai_test.go
@@ -106,7 +106,7 @@ func TestSettingModelFailsOnASafetyBlockedGeminiReply(t *testing.T) {
model := NewModel(&storepb.AISetting{Provider: storepb.AISetting_GEMINI, Endpoint: server.URL, Model: "gemini-3.5-flash", ApiKey: "test-key"})
result, err := NewReviewer(model).Review(context.Background(), &Request{Statement: "DELETE FROM banned_phrases;"}, ordersCatalog())
- require.ErrorIs(t, err, ErrEmptyReply)
+ require.ErrorIs(t, err, ErrBlockedReply)
require.Nil(t, result)
require.Equal(t, 1, requests, "a blocked turn gets no correction round")
}
diff --git a/backend/plugin/aireview/reviewer.go b/backend/plugin/aireview/reviewer.go
index a182d7d5bc..735e3effd4 100644
--- a/backend/plugin/aireview/reviewer.go
+++ b/backend/plugin/aireview/reviewer.go
@@ -19,9 +19,14 @@ import (
const (
defaultMaxModelCalls = 20
defaultTimeout = 10 * time.Minute
- // maxInvalidReplies is how many times the model is asked to correct a
- // final reply before the review fails.
+ // maxInvalidReplies is how many times the model is asked to try again, after
+ // an invalid final reply or a tool call the vendor could not parse, before
+ // the review fails.
maxInvalidReplies = 1
+ // malformedToolCallMessage answers a turn that held a tool call the vendor
+ // could not parse. That turn is not in the history, so the message has to
+ // say what happened.
+ malformedToolCallMessage = "Your last turn held a tool call that could not be parsed, so no tool ran. Call the tool again, or reply with only the JSON object from the answer format."
)
var (
@@ -29,10 +34,16 @@ var (
ErrCallLimit = errors.New("the review reached the model call limit")
// ErrDeadline means the review ran longer than its time limit.
ErrDeadline = errors.New("the review ran out of time")
- // ErrInvalidReply means the final reply was still invalid after the model was asked to correct it.
+ // ErrInvalidReply means an invalid final reply, or a tool call the vendor could not parse, came when no correction round or call was left.
ErrInvalidReply = errors.New("the model did not return a valid review result")
// ErrEmptyReply means the vendor returned neither text nor tool calls.
ErrEmptyReply = errors.New("the model returned an empty reply")
+ // ErrTruncatedReply means the reply was truncated at a token limit, so findings may be missing.
+ ErrTruncatedReply = errors.New("the model's reply was truncated at a token limit")
+ // ErrBlockedReply means the vendor blocked the request or the reply, or the model refused to review.
+ ErrBlockedReply = errors.New("the vendor blocked the review or the model refused it")
+ // ErrUnfinishedReply means the stop reason was OTHER, such as Gemini's OTHER, or a value the loop does not know.
+ ErrUnfinishedReply = errors.New("the vendor ended the turn for a reason the review does not accept")
)
// Request is one review: one sheet against one target database.
@@ -124,9 +135,34 @@ func (r *Reviewer) Review(parent context.Context, request *Request, tools Tools)
content := response.GetContent()
toolCalls := response.GetToolCalls()
- // An empty reply is a safety block, a refusal, or a garbled tool call.
- // ai.Chat carries no stop reason yet, so the loop cannot tell them apart,
- // and asking again could let a blocked review pass on the second try.
+ // A truncated reply can parse as a whole answer with findings or tool
+ // arguments missing, a blocked change was never judged, and OTHER can be
+ // a block the vendor does not name. None of the three gets a correction
+ // round, which could pass a change the model did not finish judging.
+ switch response.GetStopReason() {
+ case v1pb.AIChatResponse_TURN_COMPLETED, v1pb.AIChatResponse_STOP_REASON_UNSPECIFIED:
+ // A vendor that sends no reason, such as a compatible server without finish_reason, is judged by its reply.
+ case v1pb.AIChatResponse_TOKEN_LIMIT:
+ return nil, ErrTruncatedReply
+ case v1pb.AIChatResponse_BLOCKED:
+ return nil, ErrBlockedReply
+ case v1pb.AIChatResponse_MALFORMED_TOOL_CALL:
+ // The model meant to call a tool, so the turn is not final whatever
+ // text or tool calls came with it. None of it runs or enters the
+ // history. The turn takes a correction round from the budget of an
+ // invalid reply, and fails the review when no round or call is left.
+ if invalidReplies >= maxInvalidReplies || result.Calls >= r.maxModelCalls {
+ return nil, errors.Wrap(ErrInvalidReply, "the vendor could not parse the model's tool call")
+ }
+ invalidReplies++
+ messages = append(messages, textMessage(v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_USER, malformedToolCallMessage))
+ continue
+ default:
+ // OTHER lands here, and so does any value added later.
+ return nil, ErrUnfinishedReply
+ }
+ // An empty reply that no stop reason explains can be a block the vendor
+ // did not name, and asking again could let a blocked review pass.
if strings.TrimSpace(content) == "" && len(toolCalls) == 0 {
return nil, ErrEmptyReply
}
diff --git a/backend/plugin/aireview/reviewer_test.go b/backend/plugin/aireview/reviewer_test.go
index 05b3097dbb..102957fbec 100644
--- a/backend/plugin/aireview/reviewer_test.go
+++ b/backend/plugin/aireview/reviewer_test.go
@@ -20,6 +20,14 @@ const (
roleTool = v1pb.AIChatMessageRole_AI_CHAT_MESSAGE_ROLE_TOOL
)
+const (
+ stopTurnCompleted = v1pb.AIChatResponse_TURN_COMPLETED
+ stopTokenLimit = v1pb.AIChatResponse_TOKEN_LIMIT
+ stopBlocked = v1pb.AIChatResponse_BLOCKED
+ stopMalformedToolCall = v1pb.AIChatResponse_MALFORMED_TOOL_CALL
+ stopOther = v1pb.AIChatResponse_OTHER
+)
+
const threeLineStatement = "ALTER TABLE orders ADD COLUMN note text;\n\nDELETE FROM order_events;\n"
const validReply = `{"findings": [{"title": "Add a WHERE clause", "severity": "P0", "line": 3, "rule": "an UPDATE or DELETE that touches more rows than the author means", "evidence": "order_events has 9000000 rows", "fix": "Delete in batches by id range"}]}`
@@ -79,6 +87,11 @@ func toolCallReply(calls ...*v1pb.AIChatToolCall) modelStep {
return modelStep{response: &v1pb.AIChatResponse{ToolCalls: calls, Usage: &v1pb.AIChatUsage{TotalTokens: 100}}}
}
+func (s modelStep) stoppedFor(reason v1pb.AIChatResponse_StopReason) modelStep {
+ s.response.StopReason = reason
+ return s
+}
+
func TestReviewReturnsFindings(t *testing.T) {
t.Parallel()
@@ -184,13 +197,170 @@ func TestReviewExitPaths(t *testing.T) {
wantToolCalls: []string{"read"},
},
{
- // A safety block, a refusal, and a garbled tool call all arrive as an
- // empty reply. Asking again could let a blocked review pass.
name: "empty reply fails without a correction round",
steps: []modelStep{finalReply(" \n"), finalReply(validReply)},
wantErr: ErrEmptyReply,
wantRequests: 1,
},
+ {
+ name: "empty reply of a completed turn fails without a correction round",
+ steps: []modelStep{finalReply("").stoppedFor(stopTurnCompleted), finalReply(validReply)},
+ wantErr: ErrEmptyReply,
+ wantRequests: 1,
+ },
+ {
+ name: "empty reply with another stop reason fails without a correction round",
+ steps: []modelStep{finalReply("").stoppedFor(stopOther), finalReply(validReply)},
+ wantErr: ErrUnfinishedReply,
+ wantRequests: 1,
+ },
+ {
+ name: "reply of a completed turn is parsed",
+ steps: []modelStep{finalReply(validReply).stoppedFor(stopTurnCompleted)},
+ wantFindings: 1,
+ wantCalls: 1,
+ },
+ {
+ name: "reply without a stop reason is parsed",
+ steps: []modelStep{finalReply(validReply).stoppedFor(v1pb.AIChatResponse_STOP_REASON_UNSPECIFIED)},
+ wantFindings: 1,
+ wantCalls: 1,
+ },
+ {
+ name: "reply with another stop reason fails even when it passes",
+ steps: []modelStep{finalReply(`{"findings": []}`).stoppedFor(stopOther), finalReply(validReply)},
+ wantErr: ErrUnfinishedReply,
+ wantRequests: 1,
+ },
+ {
+ name: "turn with another stop reason runs no tool",
+ steps: []modelStep{toolCallReply(&v1pb.AIChatToolCall{Id: "call-1", Name: "read"}).stoppedFor(stopOther), finalReply(validReply)},
+ wantErr: ErrUnfinishedReply,
+ wantRequests: 1,
+ },
+ {
+ name: "reply with a stop reason added later fails",
+ steps: []modelStep{finalReply(`{"findings": []}`).stoppedFor(v1pb.AIChatResponse_StopReason(99)), finalReply(validReply)},
+ wantErr: ErrUnfinishedReply,
+ wantRequests: 1,
+ },
+ {
+ name: "truncated reply fails even when it parses",
+ steps: []modelStep{finalReply(validReply).stoppedFor(stopTokenLimit), finalReply(validReply)},
+ wantErr: ErrTruncatedReply,
+ wantRequests: 1,
+ },
+ {
+ name: "truncated reply that does not parse gets no correction round",
+ steps: []modelStep{finalReply(`{"findings": [{"title": "Add a WHERE`).stoppedFor(stopTokenLimit), finalReply(validReply)},
+ wantErr: ErrTruncatedReply,
+ wantRequests: 1,
+ },
+ {
+ name: "truncated turn with tool calls runs no tool",
+ steps: []modelStep{toolCallReply(&v1pb.AIChatToolCall{Id: "call-1", Name: "read"}).stoppedFor(stopTokenLimit), finalReply(validReply)},
+ wantErr: ErrTruncatedReply,
+ wantRequests: 1,
+ },
+ {
+ name: "blocked reply with text gets no correction round",
+ steps: []modelStep{finalReply("I can't help with that request.").stoppedFor(stopBlocked), finalReply(validReply)},
+ wantErr: ErrBlockedReply,
+ wantRequests: 1,
+ },
+ {
+ name: "blocked reply that parses still fails",
+ steps: []modelStep{finalReply(`{"findings": []}`).stoppedFor(stopBlocked)},
+ wantErr: ErrBlockedReply,
+ wantRequests: 1,
+ },
+ {
+ name: "blocked reply without text gets no correction round",
+ steps: []modelStep{finalReply("").stoppedFor(stopBlocked), finalReply(validReply)},
+ wantErr: ErrBlockedReply,
+ wantRequests: 1,
+ },
+ {
+ name: "blocked turn with tool calls runs no tool",
+ steps: []modelStep{toolCallReply(&v1pb.AIChatToolCall{Id: "call-1", Name: "read"}).stoppedFor(stopBlocked), finalReply(validReply)},
+ wantErr: ErrBlockedReply,
+ wantRequests: 1,
+ },
+ {
+ name: "tool call the vendor could not parse gets one correction round",
+ steps: []modelStep{finalReply("").stoppedFor(stopMalformedToolCall), finalReply(validReply)},
+ wantFindings: 1,
+ wantCalls: 2,
+ wantLastUserText: "Call the tool again, or reply with only the JSON object",
+ },
+ {
+ name: "model calls the tool again after the vendor could not parse the call",
+ steps: []modelStep{
+ finalReply("").stoppedFor(stopMalformedToolCall),
+ toolCallReply(&v1pb.AIChatToolCall{Id: "call-1", Name: "read"}),
+ finalReply(validReply),
+ },
+ wantFindings: 1,
+ wantCalls: 3,
+ wantToolCalls: []string{"read"},
+ },
+ {
+ name: "passing text beside an unparsed tool call gets a correction round, not a pass",
+ steps: []modelStep{finalReply(`{"findings": []}`).stoppedFor(stopMalformedToolCall), finalReply(validReply)},
+ wantFindings: 1,
+ wantCalls: 2,
+ wantLastUserText: "Call the tool again, or reply with only the JSON object",
+ },
+ {
+ name: "tool calls beside an unparsed tool call do not run",
+ steps: []modelStep{toolCallReply(&v1pb.AIChatToolCall{Id: "call-1", Name: "read"}).stoppedFor(stopMalformedToolCall), finalReply(validReply)},
+ wantFindings: 1,
+ wantCalls: 2,
+ wantToolCalls: nil,
+ },
+ {
+ name: "second tool call the vendor could not parse fails",
+ steps: []modelStep{finalReply("").stoppedFor(stopMalformedToolCall)},
+ wantErr: ErrInvalidReply,
+ wantErrText: "the vendor could not parse the model's tool call",
+ wantRequests: 2,
+ },
+ {
+ name: "second unparsed tool call with passing text fails",
+ steps: []modelStep{finalReply(`{"findings": []}`).stoppedFor(stopMalformedToolCall)},
+ wantErr: ErrInvalidReply,
+ wantErrText: "the vendor could not parse the model's tool call",
+ wantRequests: 2,
+ },
+ {
+ name: "unparsed tool call after an invalid reply has no correction round left",
+ steps: []modelStep{finalReply("not json"), finalReply("").stoppedFor(stopMalformedToolCall), finalReply(validReply)},
+ wantErr: ErrInvalidReply,
+ wantErrText: "the vendor could not parse the model's tool call",
+ wantRequests: 2,
+ },
+ {
+ name: "unparsed tool call with passing text after an invalid reply has no correction round left",
+ steps: []modelStep{finalReply("not json"), finalReply(`{"findings": []}`).stoppedFor(stopMalformedToolCall), finalReply(validReply)},
+ wantErr: ErrInvalidReply,
+ wantErrText: "the vendor could not parse the model's tool call",
+ wantRequests: 2,
+ },
+ {
+ name: "invalid reply after an unparsed tool call has no correction round left",
+ steps: []modelStep{finalReply("").stoppedFor(stopMalformedToolCall), finalReply("not json"), finalReply(validReply)},
+ wantErr: ErrInvalidReply,
+ wantErrText: "the reply is not a JSON object",
+ wantRequests: 2,
+ },
+ {
+ name: "unparsed tool call on the last permitted call fails",
+ steps: []modelStep{finalReply(`{"findings": []}`).stoppedFor(stopMalformedToolCall)},
+ maxModelCalls: 1,
+ wantErr: ErrInvalidReply,
+ wantErrText: "the vendor could not parse the model's tool call",
+ wantRequests: 1,
+ },
{
name: "model error fails",
steps: []modelStep{{err: errors.New("Gemini API returned status 401")}},
@@ -265,6 +435,55 @@ func TestReviewExitPaths(t *testing.T) {
}
}
+func TestReviewStopReasonFailureOfAnEmptyReplyIsNotErrEmptyReply(t *testing.T) {
+ t.Parallel()
+
+ tests := []struct {
+ name string
+ reason v1pb.AIChatResponse_StopReason
+ wantErr error
+ }{
+ {name: "truncated", reason: stopTokenLimit, wantErr: ErrTruncatedReply},
+ {name: "blocked", reason: stopBlocked, wantErr: ErrBlockedReply},
+ {name: "another stop reason", reason: stopOther, wantErr: ErrUnfinishedReply},
+ }
+ for _, tc := range tests {
+ t.Run(tc.name, func(t *testing.T) {
+ t.Parallel()
+
+ model := &scriptedModel{steps: []modelStep{finalReply("").stoppedFor(tc.reason)}}
+ _, err := NewReviewer(model).Review(context.Background(), &Request{Statement: threeLineStatement}, &fakeTools{})
+ require.ErrorIs(t, err, tc.wantErr)
+ require.NotErrorIs(t, err, ErrEmptyReply)
+ })
+ }
+}
+
+func TestReviewKeepsAnUnparsedToolCallTurnOutOfTheHistory(t *testing.T) {
+ t.Parallel()
+
+ tests := []struct {
+ name string
+ step modelStep
+ }{
+ {name: "empty", step: finalReply("")},
+ {name: "with text", step: finalReply(`{"findings": []}`)},
+ {name: "with tool calls", step: toolCallReply(&v1pb.AIChatToolCall{Id: "call-1", Name: "read"})},
+ }
+ for _, tc := range tests {
+ t.Run(tc.name, func(t *testing.T) {
+ t.Parallel()
+
+ model := &scriptedModel{steps: []modelStep{tc.step.stoppedFor(stopMalformedToolCall), finalReply(validReply)}}
+ _, err := NewReviewer(model).Review(context.Background(), &Request{Statement: threeLineStatement}, &fakeTools{})
+ require.NoError(t, err)
+ history := model.requests[1].Messages
+ require.Equal(t, []v1pb.AIChatMessageRole{roleSystem, roleUser, roleUser}, roles(history))
+ require.Equal(t, malformedToolCallMessage, history[2].GetContent())
+ })
+ }
+}
+
func TestReviewUnknownToolNamesTheRealTools(t *testing.T) {
t.Parallel()
diff --git a/frontend/src/types/proto-es/v1/ai_service_pb.d.ts b/frontend/src/types/proto-es/v1/ai_service_pb.d.ts
index 4ad9218bb5..1d83a88eda 100644
--- a/frontend/src/types/proto-es/v1/ai_service_pb.d.ts
+++ b/frontend/src/types/proto-es/v1/ai_service_pb.d.ts
@@ -200,6 +200,13 @@ export declare type AIChatResponse = Message<"bytebase.v1.AIChatResponse"> & {
* @generated from field: optional bytebase.v1.AIChatUsage usage = 3;
*/
usage?: AIChatUsage | undefined;
+
+ /**
+ * Why the model stopped. A truncated or blocked response can still carry content.
+ *
+ * @generated from field: bytebase.v1.AIChatResponse.StopReason stop_reason = 4;
+ */
+ stopReason: AIChatResponse_StopReason;
};
/**
@@ -208,6 +215,60 @@ export declare type AIChatResponse = Message<"bytebase.v1.AIChatResponse"> & {
*/
export declare const AIChatResponseSchema: GenMessage<AIChatResponse>;
+/**
+ * Why the model stopped generating a response.
+ *
+ * @generated from enum bytebase.v1.AIChatResponse.StopReason
+ */
+export enum AIChatResponse_StopReason {
+ /**
+ * The provider did not report why the model stopped.
+ *
+ * @generated from enum value: STOP_REASON_UNSPECIFIED = 0;
+ */
+ STOP_REASON_UNSPECIFIED = 0,
+
+ /**
+ * The model completed its turn with an answer or with tool calls.
+ *
+ * @generated from enum value: TURN_COMPLETED = 1;
+ */
+ TURN_COMPLETED = 1,
+
+ /**
+ * The response was truncated at a token limit and is incomplete.
+ *
+ * @generated from enum value: TOKEN_LIMIT = 2;
+ */
+ TOKEN_LIMIT = 2,
+
+ /**
+ * The provider blocked the request or the response, or the model refused to answer.
+ *
+ * @generated from enum value: BLOCKED = 3;
+ */
+ BLOCKED = 3,
+
+ /**
+ * The model generated a tool call that the provider could not parse.
+ *
+ * @generated from enum value: MALFORMED_TOOL_CALL = 4;
+ */
+ MALFORMED_TOOL_CALL = 4,
+
+ /**
+ * The model stopped for a reason that no other value names.
+ *
+ * @generated from enum value: OTHER = 5;
+ */
+ OTHER = 5,
+}
+
+/**
+ * Describes the enum bytebase.v1.AIChatResponse.StopReason.
+ */
+export declare const AIChatResponse_StopReasonSchema: GenEnum<AIChatResponse_StopReason>;
+
/**
* Role of a chat message.
*
diff --git a/frontend/src/types/proto-es/v1/ai_service_pb.js b/frontend/src/types/proto-es/v1/ai_service_pb.js
index be173db01f..a9f5b83a5e 100644
--- a/frontend/src/types/proto-es/v1/ai_service_pb.js
+++ b/frontend/src/types/proto-es/v1/ai_service_pb.js
@@ -10,7 +10,7 @@ import { file_v1_annotation } from "./annotation_pb";
* Describes the file v1/ai_service.proto.
*/
export const file_v1_ai_service = /*@__PURE__*/
- fileDesc("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", [file_google_api_annotations, file_v1_annotation]);
+ fileDesc("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", [file_google_api_annotations, file_v1_annotation]);
/**
* Describes the message bytebase.v1.AIChatMessage.
@@ -54,6 +54,20 @@ export const AIChatUsageSchema = /*@__PURE__*/
export const AIChatResponseSchema = /*@__PURE__*/
messageDesc(file_v1_ai_service, 5);
+/**
+ * Describes the enum bytebase.v1.AIChatResponse.StopReason.
+ */
+export const AIChatResponse_StopReasonSchema = /*@__PURE__*/
+ enumDesc(file_v1_ai_service, 5, 0);
+
+/**
+ * Why the model stopped generating a response.
+ *
+ * @generated from enum bytebase.v1.AIChatResponse.StopReason
+ */
+export const AIChatResponse_StopReason = /*@__PURE__*/
+ tsEnum(AIChatResponse_StopReasonSchema);
+
/**
* Describes the enum bytebase.v1.AIChatMessageRole.
*/
diff --git a/proto/gen/grpc-doc/openapi.yaml b/proto/gen/grpc-doc/openapi.yaml
index 3fee8a3d0f..d61500bd6b 100644
--- a/proto/gen/grpc-doc/openapi.yaml
+++ b/proto/gen/grpc-doc/openapi.yaml
@@ -16583,9 +16583,25 @@ components:
- type: "null"
title: usage
description: Token usage for this provider call, when available.
+ stopReason:
+ allOf:
+ - $ref: '#/components/schemas/bytebase.v1.AIChatResponse.StopReason'
+ title: stop_reason
+ description: Why the model stopped. A truncated or blocked response can still carry content.
title: AIChatResponse
additionalProperties: false
description: Response message for AIService.Chat.
+ bytebase.v1.AIChatResponse.StopReason:
+ type: string
+ title: StopReason
+ enum:
+ - STOP_REASON_UNSPECIFIED
+ - TURN_COMPLETED
+ - TOKEN_LIMIT
+ - BLOCKED
+ - MALFORMED_TOOL_CALL
+ - OTHER
+ description: Why the model stopped generating a response.
bytebase.v1.AIChatToolCall:
type: object
properties:
diff --git a/proto/gen/grpc-doc/v1/README.md b/proto/gen/grpc-doc/v1/README.md
index 2271ffd15c..2a9f118091 100644
--- a/proto/gen/grpc-doc/v1/README.md
+++ b/proto/gen/grpc-doc/v1/README.md
@@ -292,6 +292,7 @@
- [AIChatUsage](#bytebase-v1-AIChatUsage)
- [AIChatMessageRole](#bytebase-v1-AIChatMessageRole)
+ - [AIChatResponse.StopReason](#bytebase-v1-AIChatResponse-StopReason)
- [AIService](#bytebase-v1-AIService)
@@ -5188,6 +5189,7 @@ Response message for AIService.Chat.
| content | [string](#string) | optional | The text content of the AI response. Optional when the response only contains tool calls. |
| tool_calls | [AIChatToolCall](#bytebase-v1-AIChatToolCall) | repeated | Tool calls the AI wants to make. |
| usage | [AIChatUsage](#bytebase-v1-AIChatUsage) | optional | Token usage for this provider call, when available. |
+| stop_reason | [AIChatResponse.StopReason](#bytebase-v1-AIChatResponse-StopReason) | | Why the model stopped. A truncated or blocked response can still carry content. |
@@ -5260,6 +5262,22 @@ Role of a chat message.
| AI_CHAT_MESSAGE_ROLE_TOOL | 4 | Tool result message. |
+
+<a name="bytebase-v1-AIChatResponse-StopReason"></a>
+
+### AIChatResponse.StopReason
+Why the model stopped generating a response.
+
+| Name | Number | Description |
+| ---- | ------ | ----------- |
+| STOP_REASON_UNSPECIFIED | 0 | The provider did not report why the model stopped. |
+| TURN_COMPLETED | 1 | The model completed its turn with an answer or with tool calls. |
+| TOKEN_LIMIT | 2 | The response was truncated at a token limit and is incomplete. |
+| BLOCKED | 3 | The provider blocked the request or the response, or the model refused to answer. |
+| MALFORMED_TOOL_CALL | 4 | The model generated a tool call that the provider could not parse. |
+| OTHER | 5 | The model stopped for a reason that no other value names. |
+
+
diff --git a/proto/gen/grpc-doc/v1/index.html b/proto/gen/grpc-doc/v1/index.html
index 486ca1e81f..20312737f5 100644
--- a/proto/gen/grpc-doc/v1/index.html
+++ b/proto/gen/grpc-doc/v1/index.html
@@ -1292,6 +1292,10 @@
<a href="#bytebase.v1.AIChatMessageRole"><span class="badge">E</span>AIChatMessageRole</a>
</li>
+ <li>
+ <a href="#bytebase.v1.AIChatResponse.StopReason"><span class="badge">E</span>AIChatResponse.StopReason</a>
+ </li>
+
<li>
@@ -15097,6 +15101,13 @@ Permissions required: None (authentication required)</p></td>
<td><p>Token usage for this provider call, when available. </p></td>
</tr>
+ <tr>
+ <td>stop_reason</td>
+ <td><a href="#bytebase.v1.AIChatResponse.StopReason">AIChatResponse.StopReason</a></td>
+ <td></td>
+ <td><p>Why the model stopped. A truncated or blocked response can still carry content. </p></td>
+ </tr>
+
</tbody>
</table>
@@ -15255,6 +15266,53 @@ Frontend must echo this back unchanged when sending tool results. </p></td>
</tbody>
</table>
+ <h3 id="bytebase.v1.AIChatResponse.StopReason">AIChatResponse.StopReason</h3>
+ <p>Why the model stopped generating a response.</p>
+ <table class="enum-table">
+ <thead>
+ <tr><td>Name</td><td>Number</td><td>Description</td></tr>
+ </thead>
+ <tbody>
+
+ <tr>
+ <td>STOP_REASON_UNSPECIFIED</td>
+ <td>0</td>
+ <td><p>The provider did not report why the model stopped.</p></td>
+ </tr>
+
+ <tr>
+ <td>TURN_COMPLETED</td>
+ <td>1</td>
+ <td><p>The model completed its turn with an answer or with tool calls.</p></td>
+ </tr>
+
+ <tr>
+ <td>TOKEN_LIMIT</td>
+ <td>2</td>
+ <td><p>The response was truncated at a token limit and is incomplete.</p></td>
+ </tr>
+
+ <tr>
+ <td>BLOCKED</td>
+ <td>3</td>
+ <td><p>The provider blocked the request or the response, or the model refused to answer.</p></td>
+ </tr>
+
+ <tr>
+ <td>MALFORMED_TOOL_CALL</td>
+ <td>4</td>
+ <td><p>The model generated a tool call that the provider could not parse.</p></td>
+ </tr>
+
+ <tr>
+ <td>OTHER</td>
+ <td>5</td>
+ <td><p>The model stopped for a reason that no other value names.</p></td>
+ </tr>
+
+ </tbody>
+ </table>
+
diff --git a/proto/v1/v1/ai_service.proto b/proto/v1/v1/ai_service.proto
index 6ed21e0216..8c4ac2cb61 100644
--- a/proto/v1/v1/ai_service.proto
+++ b/proto/v1/v1/ai_service.proto
@@ -96,6 +96,22 @@ message AIChatUsage {
// Response message for AIService.Chat.
message AIChatResponse {
+ // Why the model stopped generating a response.
+ enum StopReason {
+ // The provider did not report why the model stopped.
+ STOP_REASON_UNSPECIFIED = 0;
+ // The model completed its turn with an answer or with tool calls.
+ TURN_COMPLETED = 1;
+ // The response was truncated at a token limit and is incomplete.
+ TOKEN_LIMIT = 2;
+ // The provider blocked the request or the response, or the model refused to answer.
+ BLOCKED = 3;
+ // The model generated a tool call that the provider could not parse.
+ MALFORMED_TOOL_CALL = 4;
+ // The model stopped for a reason that no other value names.
+ OTHER = 5;
+ }
+
// The text content of the AI response. Optional when the response only contains tool calls.
optional string content = 1;
@@ -104,4 +120,7 @@ message AIChatResponse {
// Token usage for this provider call, when available.
optional AIChatUsage usage = 3;
+
+ // Why the model stopped. A truncated or blocked response can still carry content.
+ StopReason stop_reason = 4;
}