seewe/sub2api
1
0
Fork 0
mirror of https://github.com/Wei-Shaw/sub2api.git synced 2026-10-16 15:43:13 +08:00
Code Issues Packages Projects Releases Wiki Activity

feat(composite): enforce reasoning effort policy

Browse Source
This commit is contained in:
Heatherm Huang
2026-08-01 21:15:44 +08:00
parent b74024c786
commit d60f4e442b
16 changed files with 250 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files
backend/internal/handler/composite_platform.go
+42
View File
@@ -1,9 +1,12 @@
package handler
import (
"strings"
"github.com/Wei-Shaw/sub2api/internal/service"
"github.com/gin-gonic/gin"
"github.com/tidwall/gjson"
)
func ensureCompositeTargetPlatform(c *gin.Context, apiKey *service.APIKey, model string) {
@@ -55,3 +58,42 @@ func effectiveAPIKeyPlatform(c *gin.Context, apiKey *service.APIKey) string {
}
return apiKey.Group.Platform
}
func openAIReasoningEffortPolicyForRequest(c *gin.Context, apiKey *service.APIKey) (string, []service.ReasoningEffortMapping, bool) {
if apiKey == nil || apiKey.Group == nil {
return "", nil, false
}
if apiKey.Group.Platform != service.PlatformOpenAI && apiKey.Group.Platform != service.PlatformComposite {
return "", nil, false
}
if effectiveAPIKeyPlatform(c, apiKey) != service.PlatformOpenAI {
return "", nil, false
}
return apiKey.Group.MaxReasoningEffort, apiKey.Group.ReasoningEffortMappings, true
}
func applyOpenAIReasoningEffortPolicyForRequest(c *gin.Context, apiKey *service.APIKey, body []byte) ([]byte, bool) {
maxEffort, mappings, ok := openAIReasoningEffortPolicyForRequest(c, apiKey)
if !ok {
return body, false
}
return service.ApplyOpenAIReasoningEffortPolicy(body, maxEffort, mappings)
}
func bindOpenAIReasoningEffortPolicyForMessagesRequest(c *gin.Context, apiKey *service.APIKey, body []byte) {
if c == nil || c.Request == nil {
return
}
// The Messages bridge synthesizes a default OpenAI effort when
// output_config.effort is omitted. Bind the group policy only for an
// explicit client value so the ceiling does not alter that default.
effort := gjson.GetBytes(body, "output_config.effort")
if !effort.Exists() || effort.Type != gjson.String || strings.TrimSpace(effort.String()) == "" {
return
}
maxEffort, mappings, ok := openAIReasoningEffortPolicyForRequest(c, apiKey)
if !ok {
return
}
c.Request = c.Request.WithContext(service.WithOpenAIReasoningEffortPolicy(c.Request.Context(), maxEffort, mappings))
}
backend/internal/handler/composite_platform_test.go
+40
View File
@@ -77,6 +77,46 @@ func TestCompositeTargetPlatformResolvedAllowsConcreteGroupWithoutResolution(t *
require.True(t, compositeTargetPlatformResolved(c, apiKey, "llama-4-maverick"))
}
func TestOpenAIReasoningEffortPolicyForCompositeTarget(t *testing.T) {
gin.SetMode(gin.TestMode)
group := &service.Group{
Platform: service.PlatformComposite,
MaxReasoningEffort: "medium",
ReasoningEffortMappings: []service.ReasoningEffortMapping{
{From: "max", To: "xhigh"},
},
}
apiKey := &service.APIKey{Group: group}
body := []byte(`{"reasoning":{"effort":"max"}}`)
openAICtx, _ := gin.CreateTestContext(httptest.NewRecorder())
openAICtx.Request = httptest.NewRequest("POST", "/v1/responses", nil)
openAICtx.Request = openAICtx.Request.WithContext(service.WithResolvedTargetPlatform(openAICtx.Request.Context(), service.PlatformOpenAI))
got, changed := applyOpenAIReasoningEffortPolicyForRequest(openAICtx, apiKey, body)
require.True(t, changed)
require.JSONEq(t, `{"reasoning":{"effort":"medium"}}`, string(got))
bindOpenAIReasoningEffortPolicyForMessagesRequest(openAICtx, apiKey, []byte(`{"output_config":{"effort":"max"}}`))
bound, changed := service.ApplyOpenAIReasoningEffortPolicyFromContext(openAICtx.Request.Context(), body)
require.True(t, changed)
require.JSONEq(t, `{"reasoning":{"effort":"medium"}}`, string(bound))
omittedCtx, _ := gin.CreateTestContext(httptest.NewRecorder())
omittedCtx.Request = httptest.NewRequest("POST", "/v1/messages", nil)
omittedCtx.Request = omittedCtx.Request.WithContext(service.WithResolvedTargetPlatform(omittedCtx.Request.Context(), service.PlatformOpenAI))
bindOpenAIReasoningEffortPolicyForMessagesRequest(omittedCtx, apiKey, []byte(`{"model":"gpt-5"}`))
omitted, changed := service.ApplyOpenAIReasoningEffortPolicyFromContext(omittedCtx.Request.Context(), body)
require.False(t, changed)
require.Equal(t, body, omitted)
grokCtx, _ := gin.CreateTestContext(httptest.NewRecorder())
grokCtx.Request = httptest.NewRequest("POST", "/v1/responses", nil)
grokCtx.Request = grokCtx.Request.WithContext(service.WithResolvedTargetPlatform(grokCtx.Request.Context(), service.PlatformGrok))
got, changed = applyOpenAIReasoningEffortPolicyForRequest(grokCtx, apiKey, body)
require.False(t, changed)
require.Equal(t, body, got)
}
func TestClientRequestedModelUsesCompositePublicModel(t *testing.T) {
gin.SetMode(gin.TestMode)
c, _ := gin.CreateTestContext(httptest.NewRecorder())
backend/internal/handler/openai_chat_completions.go
+2 -4
View File
@@ -80,10 +80,8 @@ func (h *OpenAIGatewayHandler) ChatCompletions(c *gin.Context) {
h.errorResponse(c, http.StatusBadRequest, "invalid_request_error", "Model is not supported by this OpenAI-compatible endpoint for composite groups")
return
}
if apiKey.Group != nil && apiKey.Group.Platform == service.PlatformOpenAI {
if cappedBody, changed := service.ApplyOpenAIReasoningEffortPolicy(body, apiKey.Group.MaxReasoningEffort, apiKey.Group.ReasoningEffortMappings); changed {
body = cappedBody
}
if cappedBody, changed := applyOpenAIReasoningEffortPolicyForRequest(c, apiKey, body); changed {
body = cappedBody
}
reqStream, ok := parseOpenAICompatibleStream(body)
if !ok {
backend/internal/handler/openai_gateway_handler.go
+4 -10
View File
@@ -312,10 +312,8 @@ func (h *OpenAIGatewayHandler) Responses(c *gin.Context) {
h.errorResponse(c, http.StatusBadRequest, "invalid_request_error", "Model is not supported by this OpenAI-compatible endpoint for composite groups")
return
}
if apiKey.Group != nil && apiKey.Group.Platform == service.PlatformOpenAI {
if cappedBody, changed := service.ApplyOpenAIReasoningEffortPolicy(body, apiKey.Group.MaxReasoningEffort, apiKey.Group.ReasoningEffortMappings); changed {
body = cappedBody
}
if cappedBody, changed := applyOpenAIReasoningEffortPolicyForRequest(c, apiKey, body); changed {
body = cappedBody
}
reqStream, ok := parseOpenAICompatibleStream(body)
@@ -932,6 +930,7 @@ func (h *OpenAIGatewayHandler) Messages(c *gin.Context) {
h.anthropicErrorResponse(c, http.StatusBadRequest, "invalid_request_error", "Model is not supported by this OpenAI-compatible endpoint for composite groups")
return
}
bindOpenAIReasoningEffortPolicyForMessagesRequest(c, apiKey, body)
routingModel := service.NormalizeOpenAICompatRequestedModel(reqModel)
preferredMappedModel := resolveOpenAIMessagesDispatchMappedModel(apiKey, reqModel)
reqStream := gjson.GetBytes(body, "stream").Bool()
@@ -1813,12 +1812,7 @@ func (h *OpenAIGatewayHandler) ResponsesWebSocket(c *gin.Context) {
zap.Int("candidate_count", scheduleDecision.CandidateCount),
)
maxReasoningEffort := ""
var reasoningEffortMappings []service.ReasoningEffortMapping
if apiKey.Group != nil && apiKey.Group.Platform == service.PlatformOpenAI {
maxReasoningEffort = apiKey.Group.MaxReasoningEffort
reasoningEffortMappings = apiKey.Group.ReasoningEffortMappings
}
maxReasoningEffort, reasoningEffortMappings, _ := openAIReasoningEffortPolicyForRequest(c, apiKey)
var requestPayloadHash string
// Passthrough rejects overlapping response.create frames, so one immutable
// turn-tagged slot preserves the exact mapping used for the in-flight request.