更新openai风控协议

2026-07-02 13:57:08 +08:00
parent c99c134713
commit 39b149b206
2 changed files with 134 additions and 3 deletions
@@ -28,6 +28,10 @@ var (
 	ErrAuth              = errors.New("chatgpt auth failed")
 	ErrQuotaExhausted    = errors.New("chatgpt quota exhausted")
 	ErrTemporaryUpstream = errors.New("chatgpt upstream temporary error")
+	// ErrContentPolicy marks a prompt rejected by ChatGPT's content audit. It is
+	// terminal and NOT retryable: the same prompt fails on every account, so the
+	// caller must fail fast rather than poll or fail over.
+	ErrContentPolicy = errors.New("chatgpt content policy rejection")
 )

 type Client struct {
@@ -94,7 +98,7 @@ func (c *Client) GenerateImage(ctx context.Context, accessToken, prompt, model,
 	if err != nil {
 		return nil, nil, err
 	}
-	fileIDs, sedimentIDs, err = c.pollForImage(ctx, session, accessToken, conversationID, fileIDs, sedimentIDs, 180*time.Second)
+	fileIDs, sedimentIDs, err = c.pollForImage(ctx, session, accessToken, conversationID, fileIDs, sedimentIDs, pollBudget(ctx))
 	if err != nil {
 		return nil, nil, err
 	}
@@ -791,6 +795,7 @@ func (c *Client) startImageGeneration(ctx context.Context, session tlsclient.Htt
 	defer resp.Body.Close()

 	conversationID := ""
+	asyncStarted := false
 	var fileIDs, sedimentIDs []string
 	scanner := bufio.NewScanner(resp.Body)
 	scanner.Buffer(make([]byte, 0, 1024*1024), 8*1024*1024)
@@ -816,6 +821,18 @@ func (c *Client) startImageGeneration(ctx context.Context, session tlsclient.Htt
 		newFiles, newSeds := scanForIDs(payload)
 		fileIDs = mergeStrings(fileIDs, newFiles)
 		sedimentIDs = mergeStrings(sedimentIDs, newSeds)
+		if !asyncStarted && containsAsyncMarker(payload) {
+			asyncStarted = true
+		}
+		// Async pipeline: ChatGPT no longer streams the image inline — it returns
+		// a placeholder tool turn (image_gen_async / image_gen_task_id) and
+		// delivers the asset later via conversation polling. Once we have the
+		// conversation id there is nothing more to read here, so stop instead of
+		// holding the SSE open until [DONE] (a stalled stream would otherwise burn
+		// the whole generation budget and surface as "context deadline exceeded").
+		if asyncStarted && conversationID != "" {
+			break
+		}
 	}
 	if conversationID == "" {
 		joined := strings.Join(chunks, "\n")
@@ -856,6 +873,31 @@ func (c *Client) getConversation(ctx context.Context, session tlsclient.HttpClie
 	return payload, nil
 }

+// pollBudget derives how long to poll for the async image from the caller's
+// remaining context budget, leaving headroom to resolve+download the asset
+// before the outer deadline (genCtx, 8min) fires. Async image generation under
+// load routinely exceeds the old hard-coded 180s, which surfaced as
+// "image poll timeout"; tying the budget to the deadline lets slow gens finish
+// while the context still backstops a truly stuck request.
+func pollBudget(ctx context.Context) time.Duration {
+	const (
+		maxBudget = 6 * time.Minute
+		headroom  = 25 * time.Second
+	)
+	deadline, ok := ctx.Deadline()
+	if !ok {
+		return 3 * time.Minute
+	}
+	budget := time.Until(deadline) - headroom
+	if budget < 0 {
+		budget = 0
+	}
+	if budget > maxBudget {
+		budget = maxBudget
+	}
+	return budget
+}
+
 func (c *Client) pollForImage(ctx context.Context, session tlsclient.HttpClient, accessToken, conversationID string, initialFileIDs, initialSedimentIDs []string, timeout time.Duration) ([]string, []string, error) {
 	start := time.Now()
 	fileIDs := append([]string{}, initialFileIDs...)
@@ -885,6 +927,12 @@ func (c *Client) pollForImage(ctx context.Context, session tlsclient.HttpClient,
 		newFiles, newSeds := extractImageIDs(conv)
 		fileIDs = mergeStrings(fileIDs, newFiles)
 		sedimentIDs = mergeStrings(sedimentIDs, newSeds)
+		// Fail fast on a content-audit refusal: the assistant turn carries the
+		// rejection text and no image will ever land, so polling to timeout only
+		// wastes the whole budget. Only bail while we have no asset yet.
+		if len(fileIDs) == 0 && len(sedimentIDs) == 0 && conversationRejected(conv) {
+			return nil, nil, ErrContentPolicy
+		}
 		if len(fileIDs) > 0 || len(sedimentIDs) > 0 {
 			time.Sleep(2 * time.Second)
 			conv, err = c.getConversation(ctx, session, accessToken, conversationID)
@@ -15,8 +15,8 @@ import (
 const (
 	baseURL                  = "https://chatgpt.com"
 	defaultUserAgent         = "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/149.0.0.0 Safari/537.36 Edg/149.0.0.0"
-	defaultClientVersion     = "prod-ab8a6348980a3e1d771c463b9f4f3e4e584f2769"
-	defaultClientBuildNumber = "7624276"
+	defaultClientVersion     = "prod-db390ebea64862bf1899c420a4c736e0cf639747"
+	defaultClientBuildNumber = "7904904"
 	defaultPOWScript         = "https://chatgpt.com/backend-api/sentinel/sdk.js"
 )

@@ -28,8 +28,91 @@ var (
 	scriptSrcRE          = regexp.MustCompile(`<script[^>]+src="([^"]+)"`)
 	dataBuildPathRE      = regexp.MustCompile(`c/[^/]*/_`)
 	htmlDataBuildRE      = regexp.MustCompile(`<html[^>]*data-build="([^"]*)"`)
+
+	// asyncMarkers signal that ChatGPT accepted the prompt and switched to the
+	// async image pipeline (image is delivered later via conversation polling
+	// rather than inline in the SSE stream). Their presence means "generating —
+	// keep polling", NOT failure.
+	asyncMarkers = []string{"image_gen_async", "image_gen_task_id", "trigger_async_ux"}
+
+	// contentPolicyMarkers are stable substrings of ChatGPT's content-audit
+	// refusal message. When one appears in an assistant turn the prompt was
+	// rejected upstream — no image will ever arrive, so we must fail fast
+	// instead of polling until timeout.
+	contentPolicyMarkers = []string{
+		"this request may violate our content polic",
+		"this prompt may violate our content polic",
+		"may violate our content policies",
+	}
 )

+// containsAsyncMarker reports whether the SSE payload indicates the async image
+// pipeline was engaged.
+func containsAsyncMarker(text string) bool {
+	for _, m := range asyncMarkers {
+		if strings.Contains(text, m) {
+			return true
+		}
+	}
+	return false
+}
+
+// detectContentPolicyRejection reports whether text contains a ChatGPT content
+// audit refusal. Matching is case-insensitive for the English variants.
+func detectContentPolicyRejection(text string) bool {
+	if text == "" {
+		return false
+	}
+	lower := strings.ToLower(text)
+	for _, m := range contentPolicyMarkers {
+		if strings.Contains(text, m) || strings.Contains(lower, m) {
+			return true
+		}
+	}
+	return false
+}
+
+// collectText concatenates every string found under value (recursively) into sb.
+func collectText(value any, sb *strings.Builder) {
+	switch x := value.(type) {
+	case string:
+		sb.WriteString(x)
+		sb.WriteByte('\n')
+	case map[string]any:
+		for _, item := range x {
+			collectText(item, sb)
+		}
+	case []any:
+		for _, item := range x {
+			collectText(item, sb)
+		}
+	}
+}
+
+// conversationRejected scans assistant turns of a fetched conversation for a
+// content-policy refusal.
+func conversationRejected(conversation map[string]any) bool {
+	mapping, _ := conversation["mapping"].(map[string]any)
+	for _, rawNode := range mapping {
+		node, _ := rawNode.(map[string]any)
+		message, _ := node["message"].(map[string]any)
+		if message == nil {
+			continue
+		}
+		author, _ := message["author"].(map[string]any)
+		role := strings.ToLower(strings.TrimSpace(stringValue(author["role"])))
+		if role != "assistant" {
+			continue
+		}
+		var sb strings.Builder
+		collectText(message["content"], &sb)
+		if detectContentPolicyRejection(sb.String()) {
+			return true
+		}
+	}
+	return false
+}
+
 func stringValue(v any) string {
 	switch x := v.(type) {
 	case string: