fix

apps/api/internal/processing/openai.go

@@ -243,10 +243,7 @@ type chatResponse struct {
 		CompletionTokens int `json:"completion_tokens"`
 		TotalTokens      int `json:"total_tokens"`
 	} `json:"usage"`
-	Error *struct {
-		Message string `json:"message"`
-		Type    string `json:"type"`
-	} `json:"error"`
+	Error *openAIErrorBody `json:"error"`
 }
 
 var (
@@ -466,8 +463,13 @@ func (c *OpenAIClient) doEmbed(ctx context.Context, texts []string) ([][]float32
 }
 
 func (c *OpenAIClient) doComplete(ctx context.Context, system, user string, temperature float64, maxTokens int, reasoningEffort string) (Completion, bool, error) {
+	sysRole := "system"
+	if openAIReasoningChatModel(c.Model) {
+		// GPT-5 family prefers developer over system on Chat Completions.
+		sysRole = "developer"
+	}
 	reqBody := buildChatCompletionBody(c.Model, []chatMessage{
-		{Role: "system", Content: system},
+		{Role: sysRole, Content: system},
 		{Role: "user", Content: user},
 	}, temperature, maxTokens, reasoningEffort)
 	body, err := json.Marshal(reqBody)
@@ -502,19 +504,13 @@ func (c *OpenAIClient) doComplete(ctx context.Context, system, user string, temp
 		return Completion{}, false, decErr
 	}
 	if res.StatusCode == http.StatusTooManyRequests || res.StatusCode >= 500 {
-		msg := "rate limited or server error"
-		if parsed.Error != nil && parsed.Error.Message != "" {
-			msg = TruncateError(errors.New(parsed.Error.Message))
-		}
+		msg := formatOpenAIHTTPError(res.StatusCode, parsed.Error)
 		httpErr := errors.New(msg)
 		stopWait(httpErr)
 		return Completion{}, true, httpErr
 	}
 	if res.StatusCode >= 400 {
-		msg := fmt.Sprintf("openai http %d", res.StatusCode)
-		if parsed.Error != nil && parsed.Error.Message != "" {
-			msg = TruncateError(errors.New(parsed.Error.Message))
-		}
+		msg := formatOpenAIHTTPError(res.StatusCode, parsed.Error)
 		httpErr := errors.New(msg)
 		stopWait(httpErr)
 		return Completion{}, false, httpErr

apps/api/internal/processing/openai_api_error.go

@@ -0,0 +1,110 @@
+package processing
+
+import (
+	"fmt"
+	"strings"
+)
+
+// openAIErrorBody is the error object returned by OpenAI-compatible APIs.
+type openAIErrorBody struct {
+	Message string `json:"message"`
+	Type    string `json:"type"`
+	Code    any    `json:"code"` // string or number depending on provider
+	Param   string `json:"param"`
+}
+
+func (e *openAIErrorBody) codeString() string {
+	if e == nil {
+		return ""
+	}
+	switch v := e.Code.(type) {
+	case string:
+		return strings.TrimSpace(v)
+	case float64:
+		if v == float64(int64(v)) {
+			return fmt.Sprintf("%d", int64(v))
+		}
+		return fmt.Sprintf("%g", v)
+	case int:
+		return fmt.Sprintf("%d", v)
+	default:
+		return ""
+	}
+}
+
+// formatOpenAIHTTPError turns an upstream HTTP error into a short, secret-safe
+// operator message. Intentionally avoids embedding model ids (gpt-…) so
+// TruncateError / mapPublicErrorCode do not collapse useful validation text to
+// "processing_failed".
+func formatOpenAIHTTPError(status int, errObj *openAIErrorBody) string {
+	msg := ""
+	code := ""
+	param := ""
+	typ := ""
+	if errObj != nil {
+		msg = strings.TrimSpace(errObj.Message)
+		code = errObj.codeString()
+		param = strings.TrimSpace(errObj.Param)
+		typ = strings.TrimSpace(errObj.Type)
+	}
+	lower := strings.ToLower(msg)
+	codeLower := strings.ToLower(code)
+	paramLower := strings.ToLower(param)
+
+	switch {
+	case status == 401 || codeLower == "invalid_api_key" ||
+		strings.Contains(lower, "invalid api key") ||
+		strings.Contains(lower, "incorrect api key"):
+		return "AI provider rejected the API key"
+	case status == 403:
+		return "AI provider forbidden the request"
+	case status == 429 || codeLower == "rate_limit_exceeded":
+		return "AI provider rate limited — retry later"
+	case codeLower == "insufficient_quota" || strings.Contains(lower, "insufficient_quota") ||
+		strings.Contains(lower, "exceeded your current quota"):
+		return "OpenAI quota exceeded — check billing"
+	case strings.Contains(lower, "does not exist") || strings.Contains(lower, "do not have access") ||
+		codeLower == "model_not_found":
+		return "model not found or API key lacks access — check model name and project permissions"
+	case strings.Contains(lower, "max_tokens") && strings.Contains(lower, "max_completion"):
+		return "model requires completion-token budget (legacy token cap unsupported)"
+	case paramLower == "temperature" || strings.Contains(lower, "temperature"):
+		return "model rejects custom temperature — omit temperature for reasoning chat models"
+	case paramLower == "max_tokens" || (strings.Contains(lower, "unsupported parameter") && strings.Contains(lower, "max_tokens")):
+		return "model rejects legacy token cap — use completion-token budget"
+	case paramLower == "reasoning_effort" || strings.Contains(lower, "reasoning_effort"):
+		return "model rejected reasoning effort — try none/low or omit"
+	case strings.Contains(lower, "unsupported parameter") || strings.Contains(lower, "unsupported value"):
+		if param != "" {
+			return fmt.Sprintf("unsupported chat parameter %q", param)
+		}
+		return "unsupported chat parameter for this model"
+	case status >= 500:
+		return "AI provider temporarily unavailable (server error)"
+	}
+
+	// Prefer structured fields over raw message (raw often embeds model ids / urls).
+	if code != "" && param != "" {
+		return fmt.Sprintf("provider http %d (%s param=%s)", status, code, param)
+	}
+	if code != "" {
+		return fmt.Sprintf("provider http %d (%s)", status, code)
+	}
+	if typ != "" {
+		return fmt.Sprintf("provider http %d (%s)", status, typ)
+	}
+	if status > 0 {
+		return fmt.Sprintf("provider http %d", status)
+	}
+	return "provider request failed"
+}
+
+// ProbeCompleteOptions is the Chat Completions budget for admin connection tests.
+// Explicit max_completion_tokens + reasoning_effort=none avoids GPT-5.6 defaults
+// spending the budget on hidden reasoning for a one-token "ok" reply.
+func ProbeCompleteOptions() CompleteOptions {
+	return CompleteOptions{
+		MaxTokens:       64,
+		ReasoningEffort: "none",
+	}
+}

apps/api/internal/processing/openai_api_error_test.go

@@ -0,0 +1,90 @@
+package processing
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestFormatOpenAIHTTPError_preservesUsefulHints(t *testing.T) {
+	t.Parallel()
+	cases := []struct {
+		status int
+		err    *openAIErrorBody
+		want   string
+	}{
+		{
+			status: 400,
+			err: &openAIErrorBody{
+				Message: "Unsupported parameter: 'temperature' does not support 0.2 with this model. Only the default (1) value is supported.",
+				Type:    "invalid_request_error",
+				Code:    "unsupported_value",
+				Param:   "temperature",
+			},
+			want: "model rejects custom temperature",
+		},
+		{
+			status: 400,
+			err: &openAIErrorBody{
+				Message: "Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.",
+				Param:   "max_tokens",
+			},
+			want: "model requires completion-token budget",
+		},
+		{
+			status: 404,
+			err: &openAIErrorBody{
+				Message: "The model `gpt-5.6-luna` does not exist or you do not have access to it.",
+				Code:    "model_not_found",
+			},
+			want: "model not found or API key lacks access",
+		},
+		{
+			status: 400,
+			err: &openAIErrorBody{
+				Message: "Unsupported parameter: 'max_tokens' is not supported with this model.",
+				Param:   "max_tokens",
+				Code:    "unsupported_parameter",
+			},
+			want: "model rejects legacy token cap",
+		},
+		{
+			status: 401,
+			err: &openAIErrorBody{
+				Message: "Incorrect API key provided: sk-proj-REDACTED",
+				Code:    "invalid_api_key",
+			},
+			want: "AI provider rejected the API key",
+		},
+		{
+			status: 400,
+			err: &openAIErrorBody{
+				Message: "something obscure mentioning gpt-5.6-luna",
+				Type:    "invalid_request_error",
+				Code:    "invalid_request_error",
+				Param:   "messages",
+			},
+			want: "provider http 400 (invalid_request_error param=messages)",
+		},
+	}
+	for _, tc := range cases {
+		got := formatOpenAIHTTPError(tc.status, tc.err)
+		if !strings.Contains(strings.ToLower(got), strings.ToLower(tc.want)) {
+			t.Fatalf("status=%d got=%q want contains %q", tc.status, got, tc.want)
+		}
+		if strings.Contains(got, "sk-") || strings.Contains(got, "gpt-5.6-luna") {
+			t.Fatalf("leaked secret or model id: %q", got)
+		}
+		safe := TruncateError(errString(got))
+		if safe == "processing_failed" {
+			t.Fatalf("TruncateError wiped useful hint: in=%q", got)
+		}
+	}
+}
+
+func TestProbeCompleteOptions(t *testing.T) {
+	t.Parallel()
+	opts := ProbeCompleteOptions()
+	if opts.MaxTokens != 64 || opts.ReasoningEffort != "none" {
+		t.Fatalf("%+v", opts)
+	}
+}

apps/api/internal/processing/openai_test.go

@@ -184,6 +184,14 @@ func TestOpenAIClient_doComplete_gpt56LunaUsesMaxCompletionTokens(t *testing.T)
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		var req map[string]any
 		_ = json.NewDecoder(r.Body).Decode(&req)
+		msgs, _ := req["messages"].([]any)
+		if len(msgs) < 1 {
+			t.Errorf("expected messages")
+		} else if m0, ok := msgs[0].(map[string]any); ok {
+			if m0["role"] != "developer" {
+				t.Errorf("system role=%v want developer for gpt-5.6", m0["role"])
+			}
+		}
 		if _, ok := req["temperature"]; ok {
 			t.Errorf("temperature must be omitted for gpt-5.6-luna")
 		}

apps/api/internal/processing/sanitize.go

@@ -232,17 +232,26 @@ func classifyProviderError(msg string) string {
 		strings.Contains(lower, "http 401"),
 		strings.Contains(lower, "invalid api key"),
 		strings.Contains(lower, "incorrect api key"),
-		strings.Contains(lower, "invalid_api_key"):
+		strings.Contains(lower, "invalid_api_key"),
+		strings.Contains(lower, "rejected the api key"):
 		return "AI provider rejected the API key"
 	case strings.Contains(lower, "http 403"),
-		lower == "forbidden":
+		lower == "forbidden",
+		strings.Contains(lower, "forbidden the request"):
 		return "AI provider forbidden the request"
-	case strings.Contains(lower, "http 429"),
-		strings.Contains(lower, "too many requests"),
-		lower == "rate limited":
-		return "AI provider rate limited — retry later"
 	case lower == "rate limited or server error":
 		return "AI provider temporarily unavailable (rate limited or server error)"
+	case strings.Contains(lower, "http 429"),
+		strings.Contains(lower, "too many requests"),
+		lower == "rate limited",
+		strings.Contains(lower, "rate limited —"),
+		strings.Contains(lower, "rate limited -"):
+		return "AI provider rate limited — retry later"
+	case strings.Contains(lower, "quota exceeded"):
+		return "OpenAI quota exceeded — check billing"
+	case strings.Contains(lower, "model not found"),
+		strings.Contains(lower, "lacks access"):
+		return "model not found or API key lacks access — check model name and project permissions"
 	}
 	return ""
 }