feat: read a fully failed batch's error file instead of reporting "empty output file id", so a generation job that had every request refused names the provider's actual reason in the admin panel rather than an internal guard message

internal/app/warmupcontent/batch.go

@@ -168,7 +168,7 @@ func (s *service) PollBatches(ctx context.Context) error {
 
 // pollBatchJob reconciles a single batch job.
 func (s *service) pollBatchJob(ctx context.Context, job *models.WarmupGenerationJob) error {
-	status, outputFileID, counts, err := s.gen.GetBatch(ctx, job.BatchID)
+	status, outputFileID, errorFileID, counts, err := s.gen.GetBatch(ctx, job.BatchID)
 	if err != nil {
 		return err
 	}
@@ -177,7 +177,14 @@ func (s *service) pollBatchJob(ctx context.Context, job *models.WarmupGeneration
 	switch status {
 	case "completed":
 		job.BatchOutputFileID = outputFileID
-		return s.ingestBatch(ctx, job, outputFileID, counts)
+		// A batch whose requests all failed completes with no output file; the
+		// refusals are in the error file, which has the same JSONL shape. Read
+		// it instead of failing on the empty id, or the reason is never seen.
+		resultsFileID := outputFileID
+		if resultsFileID == "" {
+			resultsFileID = errorFileID
+		}
+		return s.ingestBatch(ctx, job, resultsFileID, counts)
 	case "failed", "expired", "cancelled":
 		now := time.Now()
 		job.Status = "failed"
@@ -287,6 +294,11 @@ func (s *service) ingestBatch(ctx context.Context, job *models.WarmupGenerationJ
 		job.Status = "failed"
 		if job.Error == "" {
 			job.Error = fmt.Sprintf("all %d batch results failed", job.FailedCount)
+			// One line's reason is worth more than the count: when every
+			// request is refused it is the same reason for all of them.
+			if reason := firstResultError(results); reason != "" {
+				job.Error += ": " + reason
+			}
 		}
 	}
 	if err := s.repo.UpdateGenerationJob(ctx, job); err != nil {
@@ -355,3 +367,14 @@ func (s *service) CancelBatch(ctx context.Context, jobID uuid.UUID) error {
 	}
 	return s.repo.UpdateGenerationJob(ctx, job)
 }
+
+// firstResultError returns the first non-empty per-line error in a batch's
+// results, used to name the reason a fully failed batch produced nothing.
+func firstResultError(results []generation.BatchResult) string {
+	for i := range results {
+		if results[i].Err != "" {
+			return results[i].Err
+		}
+	}
+	return ""
+}

internal/pkg/generation/batch.go

@@ -112,19 +112,22 @@ func (c *GenerationClient) SubmitBatch(ctx context.Context, requests []BatchRequ
 	return batch.ID, file.ID, nil
 }
 
-// GetBatch returns the current status, output file ID (when completed), and
-// request counts for a batch.
-func (c *GenerationClient) GetBatch(ctx context.Context, batchID string) (status, outputFileID string, counts BatchCounts, err error) {
+// GetBatch returns the current status, the output file ID (present when at
+// least one request succeeded), the error file ID (present when at least one
+// failed), and the request counts for a batch. A batch whose requests all
+// failed still completes, with output empty and every refusal in the error
+// file, so callers must read that one to learn why.
+func (c *GenerationClient) GetBatch(ctx context.Context, batchID string) (status, outputFileID, errorFileID string, counts BatchCounts, err error) {
 	batch, err := c.client.Batches.Get(ctx, batchID)
 	if err != nil {
-		return "", "", BatchCounts{}, err
+		return "", "", "", BatchCounts{}, err
 	}
 	counts = BatchCounts{
 		Completed: int(batch.RequestCounts.Completed),
 		Failed:    int(batch.RequestCounts.Failed),
 		Total:     int(batch.RequestCounts.Total),
 	}
-	return string(batch.Status), batch.OutputFileID, counts, nil
+	return string(batch.Status), batch.OutputFileID, batch.ErrorFileID, counts, nil
 }
 
 // CancelBatch requests cancellation of an in-flight batch.
@@ -180,7 +183,12 @@ func parseBatchOutputLine(raw string) BatchResult {
 		return res
 	}
 	if line.Response.StatusCode < 200 || line.Response.StatusCode >= 300 {
+		// An error line carries {"body":{"error":{"message":...}}}, which does
+		// not fit ChatCompletion; without this the reason is reduced to a code.
 		res.Err = fmt.Sprintf("response status %d", line.Response.StatusCode)
+		if msg := batchLineErrorMessage(raw); msg != "" {
+			res.Err += ": " + msg
+		}
 		return res
 	}
 	if len(line.Response.Body.Choices) == 0 {
@@ -204,3 +212,21 @@ type namedReader struct {
 }
 
 func (n namedReader) Name() string { return n.name }
+
+// batchLineErrorMessage pulls the provider's message out of a failed batch
+// line, whose body is an error object rather than a chat completion.
+func batchLineErrorMessage(raw string) string {
+	var line struct {
+		Response struct {
+			Body struct {
+				Error struct {
+					Message string `json:"message"`
+				} `json:"error"`
+			} `json:"body"`
+		} `json:"response"`
+	}
+	if err := json.Unmarshal([]byte(raw), &line); err != nil {
+		return ""
+	}
+	return strings.TrimSpace(line.Response.Body.Error.Message)
+}