fix(llm): retry openai chat with max_completion_tokens when max_tokens is rejected

internal/llm/openai.go

@@ -43,6 +43,35 @@ func (c *openAIClient) Vision(ctx context.Context, prompt string, images []Image
 		"temperature": opts.Temperature,
 		"messages":    []map[string]any{{"role": "user", "content": content}},
 	}
+
+	b, err := json.Marshal(reqBody)
+	if err != nil {
+		return "", err
+	}
+
+	status, raw, err := c.postJSONRaw(ctx, "/v1/chat/completions", b)
+	if err != nil {
+		return "", err
+	}
+
+	// Self-heal: newer OpenAI models reject max_tokens; retry with max_completion_tokens.
+	if status == http.StatusBadRequest && bytes.Contains(raw, []byte("max_completion_tokens")) {
+		reqBody["max_completion_tokens"] = reqBody["max_tokens"]
+		delete(reqBody, "max_tokens")
+		b, err = json.Marshal(reqBody)
+		if err != nil {
+			return "", err
+		}
+		status, raw, err = c.postJSONRaw(ctx, "/v1/chat/completions", b)
+		if err != nil {
+			return "", err
+		}
+	}
+
+	if status >= 300 {
+		return "", fmt.Errorf("openai HTTP %d: %s", status, string(raw))
+	}
+
 	var out struct {
 		Choices []struct {
 			Message struct {
@@ -50,7 +79,7 @@ func (c *openAIClient) Vision(ctx context.Context, prompt string, images []Image
 			} `json:"message"`
 		} `json:"choices"`
 	}
-	if err := c.postJSON(ctx, "/v1/chat/completions", reqBody, &out); err != nil {
+	if err := json.Unmarshal(raw, &out); err != nil {
 		return "", err
 	}
 	if len(out.Choices) == 0 {
@@ -59,6 +88,7 @@ func (c *openAIClient) Vision(ctx context.Context, prompt string, images []Image
 	return out.Choices[0].Message.Content, nil
 }
 
+
 func (c *openAIClient) Transcribe(ctx context.Context, audio []byte, filename string, opts Options) (string, error) {
 	model := c.audioModel
 	if opts.Model != "" {
@@ -116,14 +146,10 @@ func (c *openAIClient) Transcribe(ctx context.Context, audio []byte, filename st
 	return out.Text, nil
 }
 
-func (c *openAIClient) postJSON(ctx context.Context, path string, body, out any) error {
-	b, err := json.Marshal(body)
-	if err != nil {
-		return err
-	}
+func (c *openAIClient) postJSONRaw(ctx context.Context, path string, b []byte) (int, []byte, error) {
 	req, err := http.NewRequestWithContext(ctx, http.MethodPost, c.base+path, bytes.NewReader(b))
 	if err != nil {
-		return err
+		return 0, nil, err
 	}
 	req.Header.Set("Content-Type", "application/json")
 	if c.key != "" {
@@ -131,15 +157,15 @@ func (c *openAIClient) postJSON(ctx context.Context, path string, body, out any)
 	}
 	resp, err := c.http.Do(req)
 	if err != nil {
-		return err
+		return 0, nil, err
 	}
 	defer func() { _ = resp.Body.Close() }()
 	raw, err := io.ReadAll(resp.Body)
 	if err != nil {
-		return err
+		return resp.StatusCode, nil, err
 	}
-	if resp.StatusCode >= 300 {
-		return fmt.Errorf("openai HTTP %d: %s", resp.StatusCode, string(raw))
-	}
-	return json.Unmarshal(raw, out)
+	return resp.StatusCode, raw, nil
 }
+
+
+

internal/llm/openai_test.go

@@ -7,6 +7,7 @@ import (
 	"net/http"
 	"net/http/httptest"
 	"strings"
+	"sync/atomic"
 	"testing"
 	"time"
 )
@@ -35,6 +36,46 @@ func TestOpenAIVision(t *testing.T) {
 		t.Fatalf("got %q err %v", got, err)
 	}
 }
+func TestOpenAIVisionMaxCompletionTokensRetry(t *testing.T) {
+	var reqCount atomic.Int32
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/chat/completions" {
+			t.Errorf("path %s", r.URL.Path)
+		}
+		reqCount.Add(1)
+		body, err := io.ReadAll(r.Body)
+		if err != nil {
+			t.Errorf("read body: %v", err)
+			return
+		}
+		bodyStr := string(body)
+		if strings.Contains(bodyStr, `"max_tokens"`) && !strings.Contains(bodyStr, `"max_completion_tokens"`) {
+			w.WriteHeader(http.StatusBadRequest)
+			_, _ = io.WriteString(w, `{"error":{"message":"Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.","type":"invalid_request_error","code":"unsupported_parameter"}}`)
+			return
+		}
+		if strings.Contains(bodyStr, `"max_completion_tokens"`) {
+			_, _ = io.WriteString(w, `{"choices":[{"message":{"content":"ok"}}]}`)
+			return
+		}
+		w.WriteHeader(http.StatusBadRequest)
+		_, _ = io.WriteString(w, `{"error":"unexpected body"}`)
+	}))
+	defer srv.Close()
+
+	c := &openAIClient{http: &http.Client{Timeout: 5 * time.Second}, base: srv.URL, key: "key", model: "gpt-5", maxTokens: 100}
+	got, err := c.Vision(context.Background(), "read", []Image{{MediaType: "image/png", Data: []byte{1}}}, Options{})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if got != "ok" {
+		t.Fatalf("got %q, want %q", got, "ok")
+	}
+	if n := reqCount.Load(); n != 2 {
+		t.Fatalf("expected 2 requests, got %d", n)
+	}
+}
+
 
 const maxMultipartMem = 1 << 20 // 1 MiB — bounded, used only in tests