Method describeImagesFallback

cli/vision_fallback.go:98–133

describeImagesFallback implements the "A" half of the hybrid vision strategy: when the active model has no native vision, it routes the attached images through a separate vision-capable model that returns a textual description, which the caller folds into the prompt so a text-only provider can still

(ctx context.Context, images []models.ImageContent)

// no vision-capable model is configured/reachable (the caller then warns and
// drops the image rather than sending something a model cannot read).
func (cli *ChatCLI) describeImagesFallback(ctx context.Context, images []models.ImageContent) (string, bool) {
	if len(images) == 0 {
		return "", false
	}

	provider, model, ok := cli.resolveVisionDescribeModel()
	if !ok {
		return "", false
	}

	client, err := cli.manager.GetClient(provider, model)
	if err != nil {
		cli.logger.Warn("vision describe-fallback: falha ao obter client",
			zap.String("provider", provider), zap.String("model", model), zap.Error(err))
		return "", false
	}

	instruction := i18n.T("vision.describe.instruction")
	userMsg := models.Message{Role: "user", Content: instruction, Images: images}

	cli.animation.UpdateMessage(i18n.T("vision.describe.in_progress", model))
	desc, err := client.SendPrompt(ctx, instruction, []models.Message{userMsg}, describeMaxTokens)
	if err != nil {
		cli.logger.Warn("vision describe-fallback: SendPrompt falhou",
			zap.String("provider", provider), zap.String("model", model), zap.Error(err))
		return "", false
	}
	desc = strings.TrimSpace(desc)
	if desc == "" {
		return "", false
	}

	// Wrap the description so the downstream (text-only) model knows this is a
	// machine-generated transcription of one or more attached images.
	return "\n\n" + i18n.T("vision.describe.note", len(images), model) + "\n" + desc + "\n", true
}

// resolveVisionDescribeModel picks the provider+model used to caption images
// for the describe-fallback. Resolution order: